Enhancing Obscured Regions in Thermal Imaging: A Novel GAN-Based Approach for Efficient Occlusion Inpainting

Abstract

This research paper presents a novel approach for occlusion inpainting in thermal images to efficiently segment and enhance obscured regions within these images. The increasing reliance on thermal imaging in fields like surveillance, security, and defense necessitates the accurate detection of obscurants such as smoke and fog. Traditional methods often struggle with these complexities, leading to the need for more advanced solutions. Our proposed methodology uses a Generative Adversarial Network (GAN) to fill occluded areas in thermal images. This process begins with an obscured region segmentation, followed by a GAN-based pixel replacement in these areas. The methodology encompasses building, training, evaluating, and optimizing the model to ensure swift real-time performance. One of the key challenges in thermal imaging is identifying effective strategies to mitigate critical information loss due to atmospheric interference. Our approach addresses this by employing sophisticated deep-learning techniques. These techniques segment, classify and inpaint these obscured regions in a patch-wise manner, allowing for more precise and accurate image restoration. We propose utilizing architectures similar to Pix2Pix and UNet networks for generative and segmentation tasks. These networks are known for their effectiveness in image-to-image translation and segmentation tasks. Our method enhances the segmentation and inpainting process by leveraging their architectural similarities. To validate our approach, we provide a quantitative analysis and performance comparison. We include a quantitative comparison between (Pix2Pix and UNet) and our combined architecture. The comparison focuses on how well each model performs in terms of accuracy and speed, highlighting the advantages of our integrated approach. This research contributes to advancing thermal imaging techniques, offering a more robust solution for dealing with obscured regions. The integration of advanced deep learning models holds the potential to significantly improve image analysis in critical applications like surveillance and security.

1. Introduction

Obscurants are substances that obscure or hide objects or areas in thermal images. Obscurants can be physical objects, such as smoke or fog, or materials explicitly designed to scatter or absorb thermal radiation. Obscurants can cause significant challenges in the processing of thermal images, as they can significantly affect the quality of the photos and reduce their usefulness for a variety of applications [1].

The most common obscurants in thermal images are smoke, fog, and clouds. Smoke can scatter and absorb thermal radiation, causing significant reductions in the visibility of objects in thermal images [2]. Fog can also scatter thermal radiation, leading to similar reductions in visibility [3]. Clouds can absorb thermal radiation and cause significant reductions in the contrast and detail of thermal images.

Several approaches have been proposed to mitigate the effects of obscurants in thermal images. One approach is to use mathematical models to correct for the scattering and absorption of thermal radiation by obscurants [1]. These models can compensate for the effects of obscurants in thermal images and restore their quality. Another approach is to use image processing techniques, such as image enhancement and restoration, to improve the quality of thermal images in the presence of obscurants [3].

The use of deep learning methods, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), has also been proposed for obscurant correction in thermal images [2]. These methods have been trained on large datasets to learn the underlying structure of thermal images and to restore the quality of thermal images in the presence of obscurants. With the increasing utility of thermal imaging in various domains, the challenges posed by obscurants in the atmosphere have become more evident. This paper proposes applying image segmentation to localize the occluded regions; later, we use GANs to inpaint the occluded segments.

This paper is organized as follows: In the next section, we provide a detailed examination of the methodology, focusing on the recent developments and techniques in the field of image segmentation and image inpainting. Section 2 outlines our methodology, including data collection strategies and the two-stage occlusion removal model architecture, which comprises mask generation using U-Net and image inpainting using Pix2Pix, followed by a discussion on training and validation. Section 3 presents our experimental results, detailing evaluation metrics such as the F1 Score and Intersection Over Union (IoU) for segmentation, Structural Similarity (SSIM), Peak Signal Noise Ratio (PSNR), and Mean Average Precision (mAP) for inpainting, along with both qualitative and quantitative results. The paper then proceeds to discussions in Section 4, future work in Section 5, and concludes with Section 6.

2. Related Work

Since our work spans image segmentation and regeneration tasks, we will highlight the most prominent and popular work developed in image segmentation, image inpainting, and generative networks for image inpainting.

2.1. Image Segmentation

The image segmentation task typically involves assigning each pixel in an image a class label, such as “sky”, “building”, “person”, etc. The goal is to achieve a dense prediction of the semantic classes in an image, providing a comprehensive understanding of the scene. In recent years, semantic segmentation has seen rapid advancements driven by deep learning-based methods. This paper will review the current state-of-the-art methods in semantic segmentation, including a discussion of recent trends and key references in the field.

One of the most important breakthroughs in semantic segmentation was the development of Fully Convolutional neural Networks (FCNs) by Long et al. [4]. FCNs were designed to address the problem of resolution loss in traditional convolutional neural networks when applied to dense prediction tasks. The authors demonstrated that an FCN trained on an end-to-end on large annotated datasets could produce state-of-the-art results in semantic segmentation.

Since then, several other deep learning-based architectures have been proposed to improve the performance of FCNs. One such architecture is U-Net [5], which introduced a symmetric encoder–decoder structure, skip connections, and upsampling layers to improve the quality of segmentation predictions. Another significant contribution was the SegNet architecture [6], which introduced a novel technique for incorporating pooling indices from the encoding stage into the decoding stage, allowing for precise localization of object boundaries.

In recent years, attention-based models have become increasingly popular in semantic segmentation. For example, the Attention U-Net [7] introduced a mechanism for weighting the features in the decoder stage based on their importance for individual prediction. Another notable attention-based model is the DenseASPP [8], which introduced a dense connection scheme for the Atrous Spatial Pyramid Pooling (ASPP) module, allowing for improved feature fusion and increased context information.

A significant trend in recent years has been the development of models that take advantage of large amounts of annotated data. For example, the DeepLab v3+ [9] introduced an Atrous Spatial Pyramid Pooling (ASPP) module that enables the network to use multi-scale contextual information. The authors demonstrated that their model, trained on a large annotated dataset, could produce state-of-the-art results in semantic segmentation. Another example is the PANet [10], which introduced a pyramid attention network to allow for efficient feature fusion and improved semantic information flow.

Thermal images are unique from traditional visual images in that they represent the temperature distribution of the objects in the scene [11] rather than their appearance. This information can detect and diagnose problems in structures, machines, or biological systems [12].

The main challenge in thermal image segmentation is that thermal images often suffer from low contrast, poor quality, and complex backgrounds. This makes it difficult to accurately segment the objects in the scene, especially when they have similar temperatures. In recent years, deep learning-based methods have emerged as a promising approach to overcome these challenges and achieve accurate segmentation results.

One of the earliest deep learning-based methods for thermal image segmentation is the U-Net by Ronneberger et al. [5]. The U-Net is an FCN that uses an encoder–decoder architecture to generate dense predictions of the semantic classes present in an image. The authors showed that the U-Net can be trained end-to-end on large annotated thermal image datasets to produce state-of-the-art results in segmentation.

Another popular approach to thermal image segmentation is to use multi-scale context information to improve the quality of segmentation predictions. For example, the DeepLab v3+ by Chen et al. [9] introduced an ASPP module that allows the network to use multi-scale contextual information. The authors demonstrated that their model, trained on a large annotated thermal image dataset, could produce state-of-the-art results in segmentation.

One breakthrough was the introduction of Adversarial Learning for Semantic Segmentation (ALSS) in 2020, which elegantly amalgamates adversarial loss with cross-entropy loss, optimizing the proficiency of semantic segmentation models [13]. The same year witnessed the unveiling of the Hybrid Attention-Aware Network (HAANet). This technique, distinct in its approach, embeds an attention mechanism designed to augment semantic segmentation’s performance during dense prediction tasks [14].

Recently, attention-based methods have also been applied to thermal image segmentation. For example, the Attention U-Net by Oktay et al. [7] introduced a mechanism for weighting the features in the decoder stage based on their importance for each prediction. Another notable attention-based model is the DenseASPP by Yang et al. [8], which introduced a dense connection scheme for the ASPP module, allowing for improved feature fusion and increased context information.

The study titled ViT-Seg: Vision Transformer for Semantic Segmentation [15] extrapolates the foundational Vision Transformer, primarily construed for image classification, to tackle segmentation dilemmas. The work underscored the potent capabilities of attention mechanisms in segmentation, highlighting the transformative capacity of such models in understanding intricate image details.

Segueing into the transformative architecture, the paper SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers by Xie et al. [16] propounded an innovative technique marrying the prowess of Transformers with conventional convolutional architectures. The fusion of these paradigms yields a promising avenue for semantic segmentation, capitalizing on the best of both worlds. Another avant-garde approach was discussed in PointRend: Image Segmentation as Rendering by Kirillov et al. [17]. Published in CVPR 2020, this research reframed segmentation as a rendering quandary, a paradigm shift that amplifies the precision of segmentation masks, especially accentuating object boundaries, which traditionally posed challenges.

Further pushing the boundaries of segmentation, the study CascadePSP: A Cascade of Global-Local-Global Pathways for Semantic Segmentation presented at AAAI 2021 by Zhang et al. [18], introduced a cascaded pathway scheme. By iteratively refining segmentation outcomes, this method sets a new benchmark in accuracy, addressing nuances previously overlooked. Last, but by no means least, the realm of medical image segmentation witnessed a monumental stride with the Swin-Unet: Unet-like Pure Transformer Architecture for Medical Image Segmentation by Chen et al. [19]. This research leverages the Swin Transformer architecture, crafting a pure transformer layout that has remarkably set a new gold standard in the domain.

2.2. Image Inpainting

Image inpainting is restoring missing or damaged parts of an image with plausible content. It is an essential problem in image processing and computer vision with many applications, including restoring old photographs, removing unwanted objects from images, and filling in missing pixels in a digital image.

Several traditional image inpainting methods have been developed over the years. These methods can be broadly classified into two categories: exemplar-based methods and diffusion-based methods.

Exemplar-based methods are the most straightforward and most intuitive image inpainting techniques. The basic idea behind exemplar-based methods is to fill in the missing parts of an image by searching for similar patches in the surrounding area and copying those patches into the missing regions. These methods are typically fast and easy to implement. However, they can lead to repetitive textures and inconsistent color in the inpainted region if the surrounding context is not carefully considered. One popular exemplar-based method is the Patch-Based Synthesis (PBS) method [20]. PBS uses a nearest-neighbor search to find the most similar patches in the surrounding area and then blends those patches to fill in the missing regions. PBS is a fast and flexible method that can handle various inpainting scenarios. However, it is limited by the quality of the available patches for synthesis.

Diffusion-based methods are another popular class of traditional image inpainting techniques. These methods use partial differential equations to propagate information from the known parts of the image into the missing regions. The basic idea is to diffuse the image content from the known parts of the image into the disappeared regions, filling in the gaps with plausible content. One popular diffusion-based method is the Navier–Stokes (NS) inpainting [21]. NS inpainting uses a partial differential equation to diffuse the image content from the known parts of the image into the missing regions. The NS equation is based on the Navier–Stokes equations from fluid dynamics, and it is used to model the flow of image content from the known parts of the image into the missing regions. NS inpainting is fast and practical but can produce unrealistic results when the image content is highly structured or non-uniform.

Another diffusion-based method is the Scale-Invariant Feature Transform (SIFT) inpainting [22]. SIFT inpainting uses the SIFT features of the surrounding area to fill in the missing regions. The basic idea is to find the SIFT features most similar to the missing areas and then use those features to fill in the gaps. SIFT inpainting is a fast and effective method, but it can be limited by the quality of the SIFT features that are available for synthesis.

2.2.1. GAN Data Generation

Developing deep neural networks has led to several new image inpainting methods that outperform traditional image inpainting techniques regarding speed, accuracy, and quality. These new methods have been based on Convolutional Neural Networks (CNNs) [23], Generative Adversarial Networks (GANs) [24], and Variational Autoencoders (VAEs) [25].

Convolutional Neural Networks (CNNs) have been the backbone of many successful DL image inpainting methods. These methods are trained on large datasets to learn images’ underlying structures and generate plausible content in the missing regions. One popular CNN-based image inpainting method is the Context Encoder [23], which uses a CNN to develop a complete image from the surrounding context and the known parts of the image. The idea behind the Context Encoder is to preserve the global structure of the image and the local details of the missing regions.

Another popular CNN-based method is Generative Adversarial Network (GAN) inpainting [24]. GAN inpainting generates plausible content in the missing regions using a generative network. The basic idea is to train a generative network to create new images similar to the training data and then use that network to fill in the missing regions of the image. GAN inpainting is a powerful method that can handle many inpainting scenarios but can be computationally expensive and challenging to train.

Variational Autoencoders (VAEs) have also been used for DL image inpainting [25]. VAEs are trained to generate new images by encoding the input image into a latent representation and then decoding that representation back into an image. One popular VAE-based method is image inpainting with deep generative models, which uses a VAE to generate plausible content in the missing regions. The basic idea is to use the VAE to learn a compact representation of the image and then use that representation to fill in the missing areas.

Generative Adversarial Networks (GANs) have been increasingly utilized in research to generate synthetic data. One such application is in the domain of tabular data, particularly for alleviating class imbalance, a common challenge in machine learning. GANs have been used to address issues where the dependent variable in a dataset exhibits class imbalance, biasing machine learning predictions. This application is significant in fields where the minority class is particularly interested, such as medical diagnosis or fraud detection [26]. Another innovative use of GANs is in the field of drug discovery. Researchers have employed GANs with adaptive training data to search for new molecular structures, significantly accelerating the discovery process and producing novel compounds with high drug-likeness scores. This method has shown promise in generating more diverse and distinct molecules than traditional approaches [27]. Furthermore, the authors in [28] used DCGAN to generate large EEG topomap images for training CNN models, aiding in cognitive load level prediction and working memory analysis. These examples underscore the versatility and potential of GANs in advancing scientific research across various disciplines.

The Pix2Pix framework, introduced by Isola et al. [29], provides a generalized approach to image-to-image translation tasks using conditional Adversarial Networks (cGANs). The primary idea is to condition the generation process on random noise (as in traditional GANs) and on a given input image.

2.2.2. Pix2Pix Key Concepts

Unlike traditional GANs that generate images from random noise, the generator in Pix2Pix produces an output image from an input image, where the goal is to transform the input image in a specific way, such as turning a sketch into a colorful image or translating satellite images to maps [29].

Conversely, the discriminator is trained to distinguish between real image pairs (input, true output) and fake image pairs (input, generated output). This conditional setup ensures that the generated images are realistic and correctly correspond to the input images [30].

To ensure the generated images are not just plausible but also close to the ground truth, an L1 loss term is introduced. This loss ensures that the individual pixel values of the generated image are close to the pixel values of the actual output image [29].

2.2.3. Relevance

The Pix2Pix framework stands out in the domain of image translation tasks, offering high-quality, context-aware translations from specific input images. It is particularly adept at photo enhancement, such as converting satellite images into detailed maps. It is a process vital for geospatial analysis and urban planning and relies on accurate map representations derived from satellite imagery [29,30]. In style transfer, Pix2Pix excels by turning sketches into realistic photos, serving as an invaluable tool in digital art and design for artists to transform their sketches into lifelike images [30,31]. Additionally, it is effective in image colorization, adding color to black and white images, which finds applications in historical photo restoration and artistic endeavors [29]. The model is also proficient in image quality enhancement, beneficial in medical imaging for clearer images leading to more accurate diagnoses, and in enhancing low-resolution images for better clarity [31]. These applications underscore Pix2Pix’s versatility and effectiveness in handling complex image translation tasks, making it a valuable asset in fields requiring precise and contextually accurate image transformations.

In our research, we picked Pix2Pix for its simple architecture and since it does not require many training samples to produce accurate results. Since the generator portion uses a U-Net architecture, we will later utilize the trained generator weights for the segmentation task. All the layers will be frozen, and the last layer will be passed to a 1 × 1 convolutional layer with a sigmoid activation function. This layer reduces the channel dimension to 1, producing a binary mask where each pixel represents the probability of belonging to the foreground class (occlusion) or background class (everything else). The sigmoid activation ensures pixel-wise values between 0 and 1, and a threshold is often applied to generate a binary mask where values above the threshold are considered object pixels, and values below are background pixels. This step is only used to optimize the model graph and reduce the number of parameters, OPs, and the total model size.

3. Methodology

3.1. Data Collection

The data collected for this study are split into two parts: the first is the occlusion part, and the second is the urban imagery mask used to train the generator. The masks applied to the images used in training the GAN are generated in the first step and explained next.

3.1.1. Occlusion Classification

The occlusion data were collected by capturing images of a pickup truck driving on a gravel road. The truck’s velocity was sufficient to kick up enough obscurants to form a dust cloud. The images were then cropped into smaller patches of size 128×128 pixels. A filter was applied to every patch where the severity of the occlusion is determined by taking the Nth derivative of the path and then subtracting the Nth derivative from the first. The severity of the occlusion is proportional to the magnitude of the difference, and the optimal value of N is determined empirically. Larger magnitudes indicate more image detail, indicating lighter image occlusion. In other words, the occlusion level is inversely related to the derivative difference. This approach is similar to the difference of Gaussian approach used to extract high-frequency information from images.

Given the Gaussian function:

$$G(x,y,\sigma )=\frac{1}{2\pi {\sigma }^2}{e}^{-\frac{x^2+y^2}{2{\sigma }^2}}$$

(1)

The magnitude of the difference between the Nth Gaussian and the first Gaussian is:

$$|G(x,y,{\sigma }_N)-G(x,y,{\sigma }_1)|$$

(2)

We distinguished between heavily occluded regions and uniform objects by calculating the average pixel value of the clear but uniform regions. For example, A clear sky has low average pixel values compared to dust clouds, and therefore, they were automatically discarded. White walls, a good example of uniform surfaces, have slightly higher temperatures; therefore, they look brighter in the image and are discarded.

We train a simple image classifier to distinguish cloudy patches from clear ones using Gray-Level Co-Occurrence Matrix (GLCM) texture analysis. More details are mentioned in [32].

We collected videos with a static camera to validate the generator on similar data. The collected images, both moving and static, have a resolution of $1920\times 1200$.

The approximate number of images captured by moving cameras is 3500, while the number of images captured by static cameras is 1000.

3.1.2. KAIST Thermal Suburban Dataset

The KAIST Multispectral Pedestrian Dataset, initially developed for pedestrian detection tasks, has also found utility in suburban image segmentation. Suburban regions, characterized by a mix of residential areas, green spaces, and infrastructure, present unique challenges in image segmentation due to varied textures, patterns, and lighting conditions. Combining high-resolution color and thermal images in the KAIST dataset offers an innovative approach to address these complexities. Thermal imaging, in particular, can highlight nuances in suburban landscapes that are otherwise subtle or invisible in RGB images, such as heat footprints of buildings or vegetation. Utilizing the KAIST dataset for suburban segmentation allows researchers and practitioners to leverage multispectral data’s complementary strengths, yielding more accurate and detailed segmentation results in suburban environments [33]. The benefit of choosing the dataset over other thermal datasets is the pairing of RGB and thermal images, which will be our way of evaluating the generative model. By running inferences with a high-performing object detection model on both the RGB and the inpainted thermal image, the accuracy will then be compared to determine the quality of the regenerated portion.

3.2. Occlusion Two-Stage Removal Model Architecture

The Occlusion Removal process consists of two parts as shown in Figure 1, starting with a segmentation model that segments the obscured parts of the image, adding a mask on those areas, and an image inpainting model that fills in the missing pixels in the image.

3.2.1. Mask Generation Using U-Net

A modified U-Net-based architecture for thermal images is trained for segmentation. The base architecture is a U-Net network. The difference is in the number of filters per layer and the optimizer function. Our variation of U-Net has significantly fewer filters and a total of 485 thousand trainable parameters compared to 40.4 million parameters in the standard U-Net. In addition, we noticed a significant increase in the performance when using root mean square propagation as the optimizer compared to the Adam optimizer. We attribute the enhanced performance to some characteristics of thermal images, such as having different noise characteristics and textures compared to RGB images. RMSprop can handle noisy gradients better as it adjusts the learning rate for each weight based on the recent history of the gradients. If a feature in the image produces a noisy gradient, then RMSprop can adaptively reduce the learning rate for such features. In addition, RMSprop’s adaptive learning rate proves significantly more effective than a fixed rate. This flexibility allows it to navigate discrepancies in contrast arising from different sensors and ensures the model learns efficiently and generalizes well across diverse inputs. Moreover, some features in thermal images may be critical but less frequent than in regular RGB images because of the nature of heat signatures. RMSprop can handle the sparsity in essential features since it can adaptively tune the learning rates for these features based on their historical gradient values. Finally, there might be specific domain challenges when it comes to thermal images, such as handling absolute vs. relative temperature differences or the emissivity of a material. If the model struggles because of those challenges, the adaptive learning rate can improve the learning process.

3.2.2. Image Inpainting Using Pix2Pix

The generator in pix2pix is a U-Net-based architecture, an encoder–decoder with skip connections. The encoder progressively downsamples the input image, capturing contextual information, while the decoder upsamples the feature maps, reconstructing the image. The skip connections help preserve spatial information lost during downsampling, which is crucial for tasks like semantic segmentation.

The discriminator in pix2pix is a convolutional PatchGAN classifier. It operates on image patches, classifying each as real or fake. This design allows the discriminator to focus on high-frequency structures, effectively capturing texture and style discrepancies between generated and natural images. This local approach enables pix2pix to produce high-quality, detailed images by ensuring local realism without assessing the entire image simultaneously.

This architecture utilizes a conditional GAN framework. The generator is conditioned on an input image to produce a corresponding output image, and the discriminator is trained to distinguish between the real and generated (fake) image pairs. The training involves backpropagating the discriminator’s assessment to improve the generator iteratively.

The data are formatted as pairs of images and labels, where images result from the segmentation model and the label is the ground truth. The input data are generated by applying a mask to the ground truth. We apply a random mask, resembling a cloud formation, to obtain images with cropped-out regions similar to the image segmentation outputs. Figure 2 demonstrates the pair formation of the network.

The masking function, as proposed, is a computational procedure designed to apply a selective masking technique to a given image. It accepts three parameters: an image, which is the input image to be processed; a scale factor, which modulates the granularity of the generated noise pattern; and a sparsity factor, which is a control parameter determining the extent of masking applied to the image.

The scale factor parameter is instrumental in generating a cloud-like mask using Perlin noise, a technique commonly used in graphics for producing natural-appearing textures [34]. Adjusting this factor alters the scale of noise features, thereby influencing the texture of the resulting mask. A higher scale factor leads to larger, more pronounced features, whereas a lower value yields a finer, more detailed noise pattern.

On the other hand, the sparsity factor quantifies the proportion of the image to undergo the masking operation, with its values typically ranging between 0 and 1. This factor directly influences the threshold in converting the normalized noise mask into a binary deletion mask. The percentile of the noise distribution determines the threshold, corresponding to the sparsity factor multiplied by 100. Consequently, a larger sparsity factor results in a greater proportion of the image being masked, whereas a smaller value reduces the extent of masking.

In the final stage of the process, the binary deletion mask is applied to the original image. This operation effectively retains the pixel values from the original image where the deletion mask is 1 (i.e., above the threshold) and sets the pixel values to zero where the mask is 0 (below the threshold). The output is an image selectively masked according to the generated pattern, which can benefit applications such as data augmentation in machine learning, where varying input data can enhance model robustness [35].

Figure 3 represents the overall process for training Pix2Pix to inpaint the missing portions in the image.

3.3. Training and Validation

The training was conducted in two stages: The first stage was training the U-Net model to segment clouds using the previously generated data, with the RMSprop as the optimizer, a learning rate of 0.001, weight decay of 0.0001, and binary cross entropy as the loss function. The second step is to simulate cropped regions as mentioned in Section 3.2.2. The model was initially trained on small deletion percentages, where only a tiny portion of the image was deleted. The deletion percentage gradually increased when performance reached a plateau. The deletion percentage ranges from 10% to 70%. The validation for both stages was conducted with data not from the training group.

3.4. Evaluation Metrics

In this section, we discuss the metrics utilized to evaluate the performance of segmentation and Generative Adversarial Network (GAN) models. These metrics are pivotal in assessing the accuracy and effectiveness of the proposed model to generate realistic images and how precisely segmented the occluded regions are.

3.4.1. Segmentation Evaluation Metrics

F1 Score

Evaluation metrics such as the F1 score [36] are commonly employed to assess classification performance, a harmonic mean of precision and recall. By striking a balance between precision (the ability to identify positives correctly) and recall (the avoidance of missing true positives), the F1 score provides a comprehensive understanding of model efficacy.

The equations for precision and recall are shown below:

$$Precision=\frac{TP}{TP+FP}$$

(3)

$$Recall=\frac{TP}{TP+FN}$$

(4)

where TP denotes the actual positives that were predicted as positive points, and FP denotes the actual negatives that were predicted as positive points. FN denotes the actual positives that were predicted as negative points.

$$F1\_score=2\ast \frac{Precision\ast Recall}{Precision+Recall}$$

(5)

Intersection over Union (IoU)

Intersection over union is an evaluation metric used to evaluate the tasks of object detection and image segmentation. IoU can determine an output image by measuring the similarity between the ground truth and the predicted region for an object in the image. The IoU is the ratio between the intersection between the predicted image and the ground truth, divided by their union. The IoU can also be represented using the following equation:

$$IoU=\frac{TP}{FP+TP+FN}$$

(6)

The authors of [37] proposed directly optimizing IoU in the deep neural network for segmentation tasks, outperforming similar architectures that utilized the softmax loss function.

3.4.2. Inpainting Evaluation Metrics

Structural Similarity (SSIM)

The structural Similarity Index is commonly used to measure the similarity between two images. It is used as a metric to assess the quality of images in the context of compression, reconstruction, and denoising. The calculation of SSIM requires the original image and the reconstructed or denoised output image. SSIM considers the luminance, contrast, and structure in the comparison. SSIM is calculated as the following:

$$SSIM(x,y)=\frac{(2{\mu }_x{\mu }_y+C_1)(2{\sigma }_{xy}+C_2)}{({\mu }_x^2+{\mu }_y^2+C_1)({\sigma }_x^2+{\sigma }_y^2+C_2)}$$

(7)

where:

• x and y are the two images being compared.
• ${\mu }_x$ and ${\mu }_y$ are the average pixel values of images x and y, respectively.
• ${\sigma }_x$ and ${\sigma }_y$ are the standard deviations of pixel values of images x and y, respectively.
• ${\sigma }_{xy}$ is the covariance of pixel values of images x and y.
• $C_1$ and $C_2$ are constants used to stabilize the division with a weak denominator, defined as $C_1={\left(k_{1}L\right)}^2$ and $C_2={\left(k_{2}L\right)}^2$, where L is the dynamic range of pixel values, which in our case is $2^7$, and $k_1=0.01$ and $k_2=0.03$ are default values.

SSIM values range from decimal −1 to 1, where 1 means identical image and lower SSIM means the data diverge from an identical set.

Peak Signal to Noise Ratio (PSNR)

Like SSIM, PSNR is a standard metric for evaluating image reconstruction quality in image compression applications. The formula for calculating PSNR is given as the following:

$$PSNR=10·{log}_{10}\left(\frac{MA{X}_I^2}{MSE}\right)$$

(8)

where:

• $MA{X}_I$ is the maximum possible pixel value of the image. For instance, if the image has 8 bits per pixel, $MA{X}_I$ is 255.
• $MSE$ stands for Mean Squared Error, which is calculated as:

$$MSE=\frac{1}{mn}\sum _{i=0}^{m-1}\sum _{j=0}^{n-1}{[I(i,j)-K(i,j)]}^2$$

(9)

where:

• $I(i,j)$ is the original image.
• $K(i,j)$ is the reconstructed (compressed or distorted) image.
• m and n are the dimensions of the images (height and width).

Figure 4 shows a sample from the evaluation set for our approach. The input and reference image pairs were used to fine-tune the model on this data for one epoch since the data size is small. Using the metrics above, we used a reference image to compare against the inpainted image.

4. Experimental Results

This section comprehensively evaluates the segmentation and GAN models employed in this study. This evaluation is structured into distinct subsections encompassing qualitative and quantitative analyses. The ‘Qualitative Results’ and ‘Quantitative Results’ subsections provide an in-depth analysis of the models’ outputs from subjective and objective standpoints.

4.1. Qualitative

In this section, we visually inspect segmented images to corroborate these findings. The images reveal clear distinctions between occluded and non-occluded areas with minimal miss-classification. This level of segmentation accuracy is particularly noteworthy, considering the inherent challenges in thermal image processing. The results undeniably showcase the efficacy of our model in accurately segmenting obscured regions, which is a critical step towards effective inpainting and enhanced image clarity in thermal imaging applications.

Figure 5a shows a sample result of the segmentation model applied to new test data taken using a different camera and in a new environment. This is a demonstration of how generalized the model is. The predicted mask corresponds to the regions where the model cannot find pixels with meaningful information, which is the region blocked by the dust cloud, which is the reason the model tends to mistake the areas with uniform intensity as a cloud if the area of the uniform surface is vast. This deficiency is shown in Figure 5b; in the regions with light occlusion with the sky as the background, the model will still confuse that region for an occluded region. It is worth noting that this confusion mostly happens in corner cases and not in the middle of the frame. We can enhance the performance in this aspect by using a larger dataset and fine-tuning the model to fit a better and more generalized model.

Figure 6a–c shows the results from the model trained for three epochs for every masking level. Given the very short training period, the model showed promising results, even for substantial deleted portions.

We used the inversion of the masks generated from the segmentation model similar to what was shown previously and fed them as input to the generator (Figure 7).

As shown in Figure 7, the model can ‘see’ the edges of the dirt road and inpaint a more visible version of the road. As expected, the model fails to predict hidden objects behind the cloud without any visible information. Therefore, the resultant inpainted image does depict reality. In the future work section, we will discuss how to improve the overall performance of both models and the entire process.

4.2. Quantitative

In

Table 1. Comparison in segmentation performance for UNet and Pix2Pix-base.

Image	IoU	F1 Score	Accuracy	Precision	Recall	Time (ms)
Reduced UNet	0.66	0.80	0.87	0.76	0.71	25
Pix2Pix Backbone	0.75	0.85	0.92	0.88	0.83	180

, we compare the results from performing cloud segmentation using the standalone reduced version of UNet with the results from using the base for the Pix2Pix generator. Predictably, the Pix2Pix generator base performance for segmentation is significantly better than the reduced version of UNet. This is due to the significant difference in the number of parameters between the two models. Pix2Pix has notably more layers, larger input sizes, and more parameters than the standalone architecture. The improved performance comes at the cost of runtime, where the reduced version takes approximately 25 ms for a single frame, and Pix2Pix takes six times as long. This is our motivation for training the reduced version for segmentation only.

In

Table 2. Table comparing the performance of our approach with common denoising and de-hazing algorithms. BM3D and BM4d are commonly used to denoise images using a stack of patches from the image and a temporal set of images. Dark Channel Prior is used to remove haze from RGB images. Transmission Map Refinement improves the estimation of the amount of light that reaches the camera without scattering, enhancing the visibility of objects in images.

Method	PSNR (dB)	SSIM	Run Time (s)
BM3D	20	0.75	2
BM4D	22	0.78	3
Dark Channel Prior	25	0.85	5
Atmospheric Light Estimation	23	0.80	4
Transmission Map Refinement	28	0.90	6
Pix2Pix on KAIST	47.3	0.90	0.2
Pix2Pix on our dataset	35.7	0.89	0.2
Our Approach	32.5	0.89	0.3

, we evaluate our approach in three stages: First, we evaluate Pix2Pix to inpaint images from the KAIST dataset with synthetic masks applied to the images. Then, we re-evaluate Pix2Pix to inpaint images from our dataset by applying random masks on clear images. Finally, we evaluate the performance when we use the same base to perform the segmentation and the inpainting. We evaluate BM3D and BM4D as a reference point of comparison.

We selected BM3D, BM4D, Atmospheric Light Estimation, and Transmission Map Refinement, as these are the most commonly used methods in dehazing and denoising applications. In our evaluation, the standalone Pix2Pix model achieved the highest PSNR performance on the KAIST dataset due to its specialization in inpainting tasks. However, its performance decreased when applied to our dataset in both cases: the random masks and the segmented occlusion. The decrease in performance is due to the segmentation side of the model not classifying all hazy pixels of the image as occluded. Therefore, the image will have some hazy parts, which will cause a drop in the generated image PSNR. Despite this drop, Pix2Pix still slightly outperformed our approach because our model is designed for segmentation and inpainting. This expected accuracy degradation in the generalized model is typical. The generalized model will experience a degradation in accuracy, as expected. Similar expectations are for the runtime of these models, as the larger models will take longer to run compared to smaller ones, given the same accuracy. As shown in Table 2, the performance of our method can be run on the edge while not compromising the quality of the results compared to other methods. This satisfies the initial assumption made in the early stages of the study. The approaches compared in the table are expected to show promising results when dealing with slight noise or small occlusions, where, as expected, the proposed approach exceeds all five methods in denoising and inpainting.

In our evaluation of the three-stage approach, as described in Table 2, we tested its performance on frames where the camera is static and the source of the dust cloud is moving. Our model showed decreased performance, compared to the model trained on KAIST, due to the higher resolution of the testing images, which were $1920\times 1200$, as opposed to the KIAST images of resolution $640\times 512$. Instead of downsizing the images by a factor of seven, we split the input images into $640\times 400$ and resized the patches to $256\times 256$. We split the image into sections to maintain some finer spatial details, which can be useful in the recognition stage. Using the higher resolution, in this case, caused the decrease in PSNR since the model was trained on images with lower resolution.

When training Pix2Pix on our dataset, the performance decreased to 35.7 dB compared to its performance when trained on the KAIST dataset. This drop in performance was caused by the smaller size of the dataset used for training. Regardless of the deterioration in performance compared to Pix2Pix, our model shows significant advancements in the other classical approaches regarding accuracy and run time.

5. Discussions and Future Work

The model’s notable strength is its ability to effectively address occlusion challenges in a two-stage process. This ability provides a comprehensive solution for obscured regions within thermal images. The model showcases a nuanced understanding of complex scenes by systematically segmenting and then inpainting, contributing to more accurate and visually coherent results.

However, it is essential to acknowledge potential limitations. One possible limitation to consider is the computational scalability of the occlusion two-stage method. Scalability is crucial when dealing with large datasets or real-time applications. Assessing the model’s performance across varying scales and computational resources will provide insights into its practical applicability.

Another limitation we faced while evaluating the model, the occlusion two-stage method, is that it struggles with complex occlusions, such as large dense clouds, as shown in Figure 8; when large portions of the frame are occluded, it becomes crucial to explore enhancements. The method often fails in large occluded areas needing more spatial context, resorting to unreliable guesswork. To mitigate this, integrating diffusion models trained with weighted contextual inputs could offer a more robust framework, leveraging their proficiency in generating quality images from noisy data. Developing advanced context-aware algorithms that adjust the influence of surrounding pixels based on occlusion complexity and employing hybrid techniques combining traditional patch-based methods with generative models could significantly enhance performance. Expanding training datasets to include more complex occlusion scenarios and utilizing semantic segmentation to inform the inpainting process are also viable strategies. These enhancements are expected to improve inpainted areas’ structural and textural accuracy and extend the practical applicability of inpainting technologies in high-precision fields.

Adaptability is another critical aspect under discussion. While the occlusion two-stage method effectively addresses obscured regions, its adaptability to diverse thermal imaging scenarios and environmental conditions is essential. The discussions explore how well the model generalizes across different contexts and whether further adaptations or fine-tuning may be necessary for optimal performance in varied real-world scenarios. Furthermore, the evaluation extends to the real-world utility of the model. Beyond its performance in controlled environments, the discussions consider the feasibility of deploying the occlusion two-stage method in practical applications such as surveillance, security, and defense. Understanding its efficacy in real-world scenarios, including detecting obscured regions under dynamic and unpredictable conditions, adds a critical layer to the model’s overall assessment.

Another potential avenue of improvement is integrating the occlusion two-stage method with real-time surveillance systems. This extension could involve optimizing the model for swift deployment and efficient processing in dynamic surveillance environments. Exploring real-time applications would contribute to the model’s practical utility and open avenues for promptly addressing obscured regions in security and surveillance settings. Expanding the dataset to include a broader range of obscurants is another promising direction. Diversifying the dataset to encompass various atmospheric interference, such as smoke, fog, or haze, would enhance the model’s robustness and generalization capabilities. This expansion would provide a more comprehensive understanding of the model’s performance under diverse environmental conditions, contributing to its adaptability across a broader spectrum of real-world scenarios.

6. Conclusions

This paper introduces a novel way to address the challenges associated with occlusion inpainting in thermal imaging, a critical requirement in the surveillance, security, and defense domains. Our approach integrates advanced Generative Adversarial Network (GAN) techniques, specifically architectures similar to Pix2Pix and UNet, to segment and enhance obscured areas in thermal images. Our method outperforms traditional approaches that struggle under atmospheric interference. The innovative use of patch-wise segmentation and inpainting techniques through deep learning enables more accurate and efficient image restoration. Based on our quantitative analysis, we have found that our combined architecture outperforms individual Pix2Pix and UNet models in terms of both accuracy and processing speed. This research presents a reliable solution to address obscured regions in thermal images, thereby paving the way for future advancements in image analysis. The potential implications of this work extend beyond immediate applications, suggesting a broader scope for deploying advanced deep-learning techniques in image processing. Ultimately, this study contributes significantly to the evolution of thermal imaging technology, marking a substantial step forward in enhancing security and surveillance capabilities.