CarrDet: Attention–Symmetry-Based Microscopic Carrier Detection for Wastewater Treatment

Abstract

The carrier is one of the key components used in wastewater treatment, which can enrich microorganisms at the surface to improve the amount of biomass in the reactor. Monitoring and adjusting the number of carriers is a key component for the processing efficiency of the ecosystem, which directly impacts wastewater treatment effectiveness. Therefore, carrier detection in wastewater microscopic images is important in the urban domestic wastewater treatment process. The current process to detect carriers is operator-dependent, which is time-consuming and expensive. Though a number of general object or cell detection approaches are used for this task, their effectiveness is limited because the carrier and background are similar and there are defective carriers as background noise. In this paper, we propose CarrDet, the first deep learning-based carrier detection framework for wastewater treatment. CarrDet uses a carrier feature block attention module and a symmetry-based defective carrier detection module to detect carriers with shallow edges and reduce false positives caused by defective carriers, respectively. To evaluate CarrDet, we propose a carrier dataset of 600 wastewater microscopic images, manually annotated by experts. Compared with state-of-the-art object detection methods, CarrDet shows superior performance in terms of both accuracy and speed, achieving a mAP of 94.32 and an IPS of 4.93. We employed CarrDet to confirm the detection results of 621 wastewater microscopic images, which were detected by inexperienced engineers who are new to the field. CarrDet added 398 unrecognized carriers with shallow edges and corrected 273 incorrect manual annotations in 5 min, which emphasizes the efficiency and practicality of CarrDet for practical business scenarios.

1. Introduction

A carrier is typically an attachable growth structure that can enrich microorganisms, such as cells and bacteria, at the surface to increase the amount of biomass in a bioreactor [1]. Adjusting and controlling the number of carriers can maintain an appropriate and effective microbial living environment [2]. The microbial system plays a vital role in wastewater treatment (WWT), as it degrades, transforms, and eliminates organic pollutants from wastewater. The accurate detection and quantification of carriers in wastewater maximizes the processing efficiency of the microbial system [3] in WWT and ensures that the effluent water quality complies with the set standards [2]. Carrier detection in microscopic images is a significant step in urban domestic WWT [4]. Manual detection methods are known to be time-consuming and subjective [5]. As a result, there has been considerable interest in deep learning-based approaches for carrier detection in wastewater microscopic images.

Though there is no method proposed for carrier detection specifically, this task can be approached by general object detection and cell detection methods. General object detection methods, e.g., YOLO [6], SSD [7], and RCNN [8], can be fine-tuned for carrier detection, using the knowledge gained from general object detection tasks. Cell detection methods [9,10] have refined the baseline, such as ResNet [11], with domain-specific knowledge from cell microscopic images.

However, because of the morphological features of carrier microscopic images, the performance of these methods is limited when they are applied to detect carriers in wastewater microscopic images directly. Carriers closely resemble the background in color, which leads to unrecognized carriers with shallow edges (Figure 1a). Moreover, there is a significant presence of defective carriers acting as background noise, which leads to false positives caused by defective carriers (Figure 1b). Inaccurate carrier detection results can lead to incorrect assessments of carrier quantity by operators, which in turn causes erroneous judgments about the current wastewater conditions. This subsequently leads to inappropriate adjustments in operational procedures, ultimately affecting the effectiveness of water purification. As a result, there is an urgent need for a dedicated approach that can effectively address these challenges in order to enhance the accuracy of carrier detection in microscopic images. This will assist operators in making correct adjustments to operations, ultimately leading to the improvement of the WWT process.

In order to tackle these challenges, we propose Attention-Symmetry-based Microscopic Carrier Detection (CarrDet), the first carrier detector of wastewater microscopic images for WWT. CarrDet mainly contains two modules: a carrier feature block attention module (CFBA) and a symmetry-based defective carrier detection module (SDCD). Specifically, we first employ the feature extractor to extract the carrier feature from the microscopic images. Subsequently, CFBA enhances the global perception of the feature through the integration of channel attention and spatial attention mechanisms. This module enlarges the contrast between the background and carrier edges and emphasizes edge features. Consequently, it effectively distinguishes unrecognized carriers with shallow edges from the background, thereby facilitating detection of them. Lastly, SDCD capitalizes on the symmetry of circular carriers to optimize the candidate frames produced by the detection head. By flipping the detection results and evaluating the pairwise similarity of corresponding positions, this approach checks if the results are symmetrical and circular. Thus, this approach reduces false positives caused by defective carriers, effectively filtering out background noise present in the images.

The main contributions of this paper are summarized as follows:

• We are the first to propose a carrier detection method specifically for wastewater microscopic images.
• We design a carrier feature block attention module and a symmetry-based defective carrier detection module to effectively reduce unrecognized results and false positives of carrier detection in wastewater microscopic images.
• We propose a carrier dataset called Carrier600 with high-accuracy annotations for evaluation. The experimental results demonstrate the effectiveness of our proposed framework and its practical applicability in real industry scenarios.

The rest of this paper is organized as follows: Section 2 provides background knowledge of carriers for wastewater treatment and related work of object detection in microscopic images. Section 3 presents the framework of CarrDet and details each component. In Section 4, we introduce the dataset, experimental settings, compared SOTA methods, and experimental results. Section 5 presents the ablation study and discusses the impact of carriers similarity threshold $\theta$. We conducted CarrDet in industry practice and show the effectiveness in Section 6. Finally, we draw conclusions in Section 7.

2. Related Work

2.1. Carrier for Wastewater Treatment

The carrier is typically a material with a large specific surface area and favorable bioadhesive properties [12], such as high concentration powder carriers. These carriers can enrich microorganisms at the surface to improve the amount of biomass in the bioreactor [1]. Adding carriers into wastewater leads to an increased concentration of mixed liquor suspended solids [13] and establishes a symbiotic microbial “double sludge” ecosystem comprising suspended growth and attached growth [14]. This ecosystem not only enhances the shock resistance of the WWT system but also improves the loading capacity of pollutants and settling performance of sludge [13]. This system plays a vital role in WWT, as it transforms and degrades organic pollutants from wastewater to meet effluent standards. Moreover, the carrier contains specialized organic alternative carbon sources, like polyhydroxyalkanoates (PHA), which can enhance the biological denitrification process and can substantially improve denitrification efficiency, thereby achieving biological purification of the wastewater [15]. By regulating carrier quantity, an optimal microbial living environment can be sustained, maximizing microbial system efficiency in WWT and ensuring effluent water quality adherence to established standards [2]. Hence, in the urban domestic WWT process, the recognition and quantification of carriers are of paramount importance. While image analysis methods have been used by wastewater treatment plants for the last many years [16], there is still a lack of specialized carrier detection methods in wastewater microscopic images.

2.2. Object Detection in Microscopic Images

Object detection in microscopic images is a common task in various biological and medical applications. Common methods include two categories, fine-tuned general object detection and domain-specific cell detection.

General object detection methods can be broadly categorized into one-stage and two-stage detectors, each focusing on different aspects. One-stage detectors such as single-shot detection (SSD) [7] and you only look once (YOLO) prioritize inference speed, while two-stage detectors such as the regional convolutional neural network (R-CNN) family, including R-CNN [8], Fast R-CNN [17], Faster R-CNN [18], and Mask R-CNN [19], prioritize detection accuracy.

Detection of certain types of cells in microscopy images is of significant interest to a wide range of biology and clinical practices [20]. To effectively detect cells in microscopic images, a range of detection methods have been developed, each suited to different cells. Prangemeier et al. [21] presented an attention-based cell detection transformer for yeast cell detection. Jiang et al. [22] verified key points and centroid points using the unique geometric sequence of cells for mammalian cell detection. Mao et al. [23] presented an optimized training algorithm for circulating tumor cell detection. Ji et al. [10] integrated upsampled multiscale features from FPN into a single layer for cell detection on IHC-stained images. Sun et al. [24] incorporated the embedding layer into the RPN to learn discriminative features based on similarity for histological cell detection. Abousamraet et al. [25] applied representation learning and deep clustering techniques to learn the morphological features of cells and representation of spatial context for breast cancer, colorectal cancer, and lung cancer cell detection. Shakarami et al. [26] used depthwise separable convolution to reduce detector parameters and distance intersection over union for blood cell detection. Liu et al. [9] refined the detection results using shape constraints of images for CD56 detection.

However, the direct application of these methods to wastewater microscopic images may lead to false positives caused by defective carriers and unrecognized carriers with shallow edges, due to the morphological specificity of carriers. Therefore, it is imperative to develop an object detection method specifically designed for carriers in wastewater microscopic images.

3. Framework

The overall network architecture of CarrDet, as depicted in Figure 2, consists of four main steps: the feature extractor, carrier feature block attention module, detection head, and symmetry-based defective carrier detection module. From the input image, the feature extractor learns feature representation. Then, we strengthen the attention on the edge of the carrier through the channel attention and spatial attention mechanism of the carrier feature block attention module. With these features as input, the detection head generates a series of carrier candidate frames. Following that, the symmetry-based defective carrier detection module eliminates defective carrier candidate frames by pairwise similarity evaluation of corresponding positions. Finally, the refined candidate frames are concentrated to form the final output. The detailed implementations of each module are illustrated as follows.

3.1. Feature Extractor

We construct a feature extractor using ResNet50 [11] and FPN [27] to extract features f from the initial input image x. The process of feature generation is as follows:

$$f=F\left(x\right)$$

(1)

Here, $f\in {\mathbb{R}}^{H\times W\times C}$, and $(H,W,C)$ represents the height, the width, and the number of channels for the extracted feature f. Specifically, we utilize the features extracted from the top three stages of ResNet50 and subsequently feed these features into FPN to achieve effective multiscale feature fusion.

3.2. Carrier Feature Block Attention

Due to subtle distinctions between the carrier edge and the background, the accurate recognition of carrier positions becomes a challenging task. To address this issue, we introduce the carrier feature block attention module, which comprises two essential components: channel attention and sparse attention. The channel attention emphasizes the “what” aspect, and the sparse attention prioritizes the “where” aspect [28]. This enhancement aims to focus on information related to carriers with shallow edges. By incorporating these attention mechanisms, the network can more accurately detect carriers in challenging scenarios involving color similarity. The following describes the details of each attention module.

3.2.1. Channel Attention

We flatten the feature f along the channel axis by using average pooling operations in order to aggregate spatial information, thereby efficiently computing the channel attention. The process can be summarized as

$$F_c=AvgPool\left(f\right)$$

(2)

where $F_c\in R^{C\times 1\times 1}$ denotes the channel compression feature. Then the similarity measures are performed to evaluate the cosine similarity between any pair of channels, which leads to the derivation of the similarity matrix $S_c$ through cosine similarity calculations:

$$S_c=\frac{F_c^i·F_c^j}{∥F_c^i∥\times ∥F_c^j∥}$$

(3)

where $F_c^i$, $F_c^j$ are the compression feature of the i-th and j-th channel of f, with $i,j\in \left[1,C\right]$ and $S_c\in R^{C\times C}$.

To determine how much attention should be given to each channel during further processing, combine the compression features of the channels $F_c$ and the similarity matrix $S_c$ in a matrix multiplication operation, resulting in a channel attention vector $A_c\in R^{C\times 1\times 1}$. $A_c$ represents the importance or relevance of each channel in the feature maps. Finally, feature f along the channel axis is multiplied with the channel attention vector $A_c$, generating image feature $F_{channel}\in R^{H\times W\times C}$.

This operation enhances the contrast between carrier edge channel features and background channel features, emphasizing the former while suppressing the influence of the latter.

3.2.2. Sparse Attention

In comparison to channel attention, our proposed sparse attention module emphasizes the informative spatial locations “where” in a complementary manner.

We perform spatial segmentation on $F_{channel}$ by merging the first and second dimensions of the image features together, which effectively reduces the dimension, and obtain a feature vector $F_p\in R^{HW\times C}$. Subsequently, we employ cosine similarity to assess the similarity between different spatial positions within the image, yielding a similarity matrix $S_s$. $S_s$ represents the pairwise similarities between various spatial positions. The calculation formula is presented below:

$$S_s=\frac{F_p^i·F_p^j}{∥F_p^i∥\times ∥F_p^j∥}$$

(4)

where $F_p^i$, $F_p^j$, respectively, denote the feature vectors corresponding to the i-th and j-th spatial position within the image, with $i,j\in \left[1,HW\right]$, and $S_s\in R^{HW\times HW}$. Then, the feature $F_p$ is subjected to matrix multiplication with the similarity matrix $S_s$, resulting in the spatial attention vectors $A_s\in R^{HW\times C}$. We obtain $F_{sparse}$ by multiplying $F_{channel}$ and $A_s$, $F_{sparse}\in R^{H\times W\times C}$.

This module aims to enhance the attention given to various spatial positions of carriers in the image, which leads to improved recognition of the edge’s spatial features, while potentially minimizing the impact of background.

3.3. Detection Head

Given image features $F_{sparse}$ as input, the detection head predicts the carrier’s position information within the image, represented by $(x,y,h,w)$. Here, $(x,y)$ denotes the coordinates of the carrier’s center point, while $(h,w)$ represents the height and width of the carrier. By leveraging the acquired carrier candidate frames, the image can be effectively partitioned, allowing for the extraction of localized carrier image features $F_{carrier}\in R^{h\times w\times c}$.

3.4. Symmetry-based Defective Carrier Detection

To address the challenge of the false positives caused by defective carriers, it is essential to consider the morphological differences between defective and normal carriers. Thus, we combine the carrier morphological features into our method. The intuitive idea is two-fold: (a) The carrier is circular in wastewater microscopic images. (b) The normal carrier is symmetrical along both its horizontal and its vertical midline. Based on these two observations, we develop a novel module suitable for reducing the false positives caused by defective carriers by evaluating the similarity between the corresponding positions of carriers, described below.

Since the carriers are all circular in shape, we can equally divide each of them predicted by the detection head into n blocks, thereby obtaining localized carrier image block features $F_b\in R^{\frac{h}{n}\times \frac{w}{n}\times c}$.

Then we evaluate the similarity of symmetrically paired image blocks, respectively. This process generates similarity matrix $S_{symmetry}\in R^{n\times n}$, which can be mathematically expressed as follows:

$$S_{symmetry}=\frac{F_b^i·F_b^{is}}{∥F_b^i∥·∥F_b^{is}∥}$$

(5)

where $F_b^i$ represents the feature vectors of the i-th image block, with $i\in \left[1,n\right]$ . Furthermore, $F_b^{is}$ represents the symmetrically positioned image block of $F_b^i$ after undergoing flipping.

Subsequently, in order to intuitively and conveniently determine whether the image block is symmetrical, we construct a matrix I for recognizing based on $S_{Symmetry}$. The specific calculation formula is as follows:

$$I=\left\{\begin{array}{c}1,if{S}_{symmetry}^{ij}>\theta \hfill \\ 0,if{S}_{symmetry}^{ij}<\theta \hfill \end{array}\right.$$

(6)

where $S_{symmetry}^{ij}$ represents the element in the row i and column j of $S_{symmetry}$, and $\theta$ corresponds to the preset carrier image blocks similarity threshold. If all elements in the matrix I approach 1 very closely, it indicates that the carrier is normal; therefore, we should retain its position information$(x,y,h,w)$. Conversely, the presence of any element in matrix I that significantly approaches 0 indicates that the observed carrier is not a symmetrical circle, which is a false positive caused by a defective carrier. By removing candidate frames containing such defective carriers, we can effectively refine the detection results and enhance the overall recognition accuracy.

4. Experiment

4.1. Dataset

Within the WWT process, carrier detection was performed in three stages: detection of pure particulate carriers, carrier detection in microscopic images of sewage mixtures, and carrier detection in microscopic images of sludge. At present, we have only conducted the first stage of the relevant research. We collected 600 images, each with a size of 5440 × 3648 pixels, using an OLYMPUS CX43 biological microscope at 400× optical magnification to build our own dataset, Carrier600. It’s important to note that the dataset was provided by Sanyou Environmental Protection Technology Co., Ltd., located in Changsha, China, at wastewater treatment plants. And Carrier600 remains confidential due to its proprietary nature. The dataset was manually annotated by two experts for two weeks, and a total of 15,614 carriers were annotated. The distribution of the number of carriers in each image is shown in Figure 3. We divided the dataset according to the ratio of 8:1:1 for model training, verification, and testing.

4.2. Experimental Settings

Our experimental environment is a 32G NVIDIA TESLA V100 GPU. To ensure equity and consistency, all our experiments are carried out under the PaddleDetection platform [29]. All models are trained for 100 epochs, and the batch size is set to 8. We use Adamw [30] as the optimizer to optimize the model, and the learning rate is set to 0.0001. During the evaluation phase, the widely accepted mean average precision (mAP) metric is employed, using an intersection over union (IoU) threshold of 0.5 as the standard criterion for assessing object detection accuracy. Additionally, the images per second (IPS) metric is employed to evaluate the efficiency of the detection method.

4.3. Compared State-of-the-Art Methods

Detection methods designed for certain cell types in microscopy images are specific to proprietary datasets and consequently cannot be directly applied to Carrier600. Therefore, to verify the effectiveness of our method, we select several state-of-the-art object detection models to compare, including Sparse R-CNN [31], ConvNeXt [32], Deformable DETR [33], DINO [34], TOOD [35], and ViT [36]. Specifically, these models are initially trained on Carrier600. Sparse-RCNN provides a fixed sparse set of learned object proposals to the object recognition head for performing classification and location tasks. ConvNeXt introduces advancements in the architecture of ViT, enabling the construction of a hierarchical model. Deformable DETR utilizes attention modules that selectively focus on a small set of key sampling points surrounding a reference. DINO employs a contrastive training approach for denoising, a mixed query selection method for anchor initialization, and a look-forward-twice scheme for box prediction. TOOD addresses the challenge of aligning object classification and localization through a learning-based approach. ViT segments an image into blocks and provides a sequence of linear embeddings of these blocks as input to the transformer.

4.4. Experimental Results

Table 1. Detection results on Carrier600. “Method” refers to specific algorithmic implementations, whereas “Backbone” refers to the core network architecture in deep learning models.

Method	Backbone	mAP(%)	IPS
Sparse R-CNN [31]	ResNet50	83.59	3.68
ConvNeXt [32]	ResNet50	85.47	1.45
Deformable DETR [33]	ResNet50	86.42	3.08
DINO [34]	ResNet50	87.36	2.52
TOOD [35]	ResNet50	90.90	4.08
ViT [36]	VisionTransformer	90.91	3.47
CarrDet	ResNet50	94.32	4.93

shows the results of these methods on Carrier600. It is evident that our proposed framework CarrDet achieves the highest mAP value of 94.32%, which is 10.73% higher than Sparse R-CNN, 8.85% higher than ConvNeXt, 7.9% higher than Deformable DETR, 6.96% higher than DINO, 3.42% higher than TOOD, and 3.41% higher than ViT. Moreover, our method also performs similarly well at processing speed. CarrDet achieves the highest IPS value of 4.93, which means processing 4.93 images per second. Notably, it shows enhancements compared to other models, surpassing Sparse R-CNN by 1.25 IPS, ConvNeXt by 3.48 IPS, Deformable DETR by 1.85 IPS, DINO by 2.41 IPS, TOOD by 0.85 IPS, and ViT by 1.46 IPS.

The superiority of our proposed method CarrDet over general object detection methods can be attributed to these key factors. Firstly, general object detection methods tend to overlook carriers with shallow edges, as depicted in the red bounding boxes in Figure 4b. In contrast, CarrDet uses the CFBA module to specifically focus on carriers’ edge features, effectively distinguishing carriers with shallow edges from the background and reducing the unrecognized rate. Secondly, as shown in the blue bounding boxes in Figure 4c, general object detection methods may erroneously recognize defective carriers. To address this, CarrDet implements the SDCD module to judge whether the shape of the carrier is a complete circle. This process efficiently filters out defective carriers, which leads to a noteworthy enhancement in detection accuracy. Finally, CarrDet uses components with a relatively simple architecture as the feature extractor and detection head, resulting in a faster image processing speed.

5. Discussion

5.1. Ablation Study

We conduct ablation experiments to evaluate the effectiveness of two modules, CFBA and SDCD, and assessed the impact of carriers similarity threshold $\theta$. In all the runs, we use the same experimental setting as above.

5.1.1. Impact of CFBA

In this part, we examine the impact of CFBA, which comprises channel attention and sparse attention, on the performance of carrier detection in microscopic images. To conduct a thorough evaluation, we perform three sets of comparative experiments on Carrier600 and present the results in

Table 2. Detection results on Carrier600.

CFBA (Channel Attention)	CFBA (Channel Attention)	SDCD	mAP(%)
			84.72
✓			88.90
	✓		89.37
✓	✓		92.93
		✓	89.04
✓	✓	✓	94.32

. The experiments involve the following: (a) only utilizing channel attention for carrier detection, (b) only utilizing sparse attention for carrier detection, and (c) employing both channel attention and sparse attention for carrier detection.

The results demonstrate that using channel attention alone improves mAP by 4.18% compared to not using it. Similarly, only using sparse attention improves the mAP by 4.65% compared to not using it. Furthermore, when both channel attention and sparse attention are combined, it outperforms the two individual experiments, and its mAP reached 92.93%.

In the absence of CFBA, general object detection methods may encounter challenges in accurately distinguishing carriers with shallow edges from the background, as their colors often bear a resemblance. Consequently, the detection of such carriers may be missed. However, by leveraging both channel attention and sparse attention, CFBA achieves a more precise focus on the channel information at each pixel and the spatial locations of carriers’ edges. These complimentary attention mechanisms contribute to effective background suppression while highlighting carriers’ edge features. As a result, CFBA significantly amplifies the contrast between carriers and the background, thereby reducing unrecognized carriers with shallow edges.

5.1.2. Impact of SDCD

In a manner analogous to the experiments conducted to assess the impact of CFBA, we conduct experiments specifically to evaluate the individual impact of SDCD on carrier detection. For this purpose, we design the following experiments for the SDCD module: (c) utilizing the SDCD module alone without CFBA, and (d) incorporating both CFBA and SDCD simultaneously. The results presented in Table 2 indicate that using the SDCD module alone improves mAP by 4.32% compared to the baseline. Moreover, employing both CFBA and SDCD outperforms the experiment (c) with an improvement of 1.39%.

The poor performance of general object detection methods can be partly attributed to the presence of defective carriers, which act as background noise. However, CarrDet uses the cosine similarity of the symmetric positions of the candidate boxes to perform secondary verification on the results of the detection head. Through this process, defective carriers are accurately filtered out, which leads to a substantial reduction in false detection results.

5.2. Impact of Carriers Similarity Threshold $\theta$

In formula (6), the threshold $\theta$ plays a crucial role in assessing the level of similarity between pairwise image blocks. If $\theta$ is set too low, the defective carriers are regarded as normal carriers, while if $\theta$ is set too high, the normal carriers are regarded as defective carriers. Therefore, it is very important to choose an appropriate threshold $\theta$. The importance of threshold selection has been emphasized in various studies, highlighting its significant impact on the performance of image processing and pattern recognition tasks [37,38] . To thoroughly investigate its impact, we conduct four sets of comparative experiments on Carrier600. In each set, the cosine distance threshold $\theta$ is set to 0.6, 0.7, 0.8, and 0.9, respectively. These experiments aim to explore its influence on carrier detection performance.

As shown in Figure 5, when the cosine distance threshold $\theta$ is 0.8, the highest mAP value of 94.32% is obtained, which is 2.69%, 1.83%, and 0.53% higher than the thresholds of 0.6, 0.7, and 0.9, respectively. Based on the trend of the mAP value as the cosine distance threshold $\theta$ changes from 0.6 to 0.9, we observe that CarrDet performs best when the cosine distance threshold $\theta$ is around 0.8. The experimental results demonstrate that the performance of our proposed framework is strongly influenced by the cosine distance threshold $\theta$. This underscores the potential constraints of our framework that should be considered during practical use.

Future research could explore automated methods for threshold selection, such as using machine learning algorithms to dynamically adjust $\theta$ based on the dataset characteristics and real-time feedback. This could eliminate the need for manual threshold setting and improve detection accuracy across varied datasets.

6. Industry Practice

During the practical domestic WWT process, approximately 6000 wastewater microscopic images are automatically generated on a daily basis. Subsequently, manual carrier detection is conducted at Environmental Protection Technology Co., Ltd. to recognize and quantify carriers, which establishes the basis for subsequent steps in WWT.

However, due to constraints in terms of manual work and time costs, it is impractical to have a sufficient number of experts for the accurate recognition task. This could potentially compromise the efficiency and effectiveness of the subsequent WWT process. Therefore, the necessity of an efficient and continuously integrated automated carrier detection method becomes crucial.

To validate the applicability of our method in a practical scenario, we collected 621 wastewater microscopic images that were detected by inexperienced engineers from the wastewater treatment plant. After being examined by two experts over a week, the corrected results of CarrDet show that CarrDet successfully rectified a total of 231 images, 166 of which were accurate, while 65 images were found to have incorrect assignments. Specifically, out of the 166 accurately corrected images, 398 unrecognized carriers with shallow edges were correctly added by our method, and 273 false positives caused by defective carriers were rectified. On average, this resulted in 2.4 accurate additions and 1.6 correct rectifications per image. The results show that manual carrier detection in wastewater microscopic images has limitations, while CarrDet’s real-world performance is excellent in assisting manual detection.

7. Conclusions

We propose CarrDet, the first deep learning-based framework that is specifically designed for carrier detection in wastewater microscopic images for the WWT process. By leveraging a carrier feature block attention module and a symmetry-based defective carrier detection module, CarrDet addresses the detection challenge of carriers with shallow edges and false positives caused by defective carriers, respectively. We construct a carrier detection dataset of wastewater microscopic images, which greatly facilitates the development of carrier detection. Experimental results show the superior performance of CarrDet, achieving a mAP of 94.32 and an IPS of 4.93, to accurately recognize carriers and improve water purification processes. Our industry practice proves the effectiveness of CarrDet in practical business WWT applications. In future work, we will try to optimize the performance of carrier detection for microscopic images that are derived from different stages of WWT, such as sewage mixtures and sludge.