Editorial on the Special Issue: New Trends in Image Processing III

The image processing field is undergoing a significant transformation owing to rapid advancements in deep learning, computer vision, and artificial intelligence. This has led to a surge in innovation in image classification [1], object detection [2,3,4], object segmentation [4,5], and image super-resolution [6], a crucial aspect of this domain that enables a more precise and versatile analysis of visual data. In recent years, image super-resolution has garnered significant interest, particularly in the domain of enhancing image quality. The primary objective of image super-resolution is to upgrade low-resolution and blurry images to high-resolution images with a primary focus on single-image super-resolution. Several trends have emerged in this field, including reconstruction-based methods [7,8,9]. These techniques have numerous applications, such as improving security in urban areas, object recognition, small object detection [10], face recognition in surveillance videos [11], medical diagnostics [12,13,14], remote sensing, astronomy [15,16,17], and microscopy image processing [18,19].

Furthermore, the domain of interest has observed a notable trend towards self-supervised learning, which enables models to extract robust features and representations from vast datasets, resulting in a reduced need for labeled data. These advancements have been applied in areas such as autonomous vehicles, medical imaging, and security surveillance. Recent breakthroughs include achieving higher precision and objectivity in disease classification and tumor microenvironment descriptions [20], outperforming CNNs in image classification with a pure transformer [21], prominent applications for remote sensing imagery, such as land use/land cover classification [22], strong zero-shot classification performance based on natural language names [23], and guidelines for and evaluation of clinically explainable AI in medical image analysis [24]. However, this approach still faces challenges, particularly in the case of limited labeled samples, which is addressed through CNN-based frameworks [25], especially in hyperspectral image classification [26]. In this research, the utilization of lightweight networks as student models and Emonet as teacher models integrates knowledge distillation, resulting in an efficient facial expression detection system with a minimal decrease in performance and notable improvements in computational efficiency, making it suitable for practical applications [27]. There are significant improvements in real-time image detection with works such as the state-of-the-art YOLO framework, which can be implemented in real-time on edge computing platforms [28], recent innovations in image classification approaches have seen the incorporation of quantum computing, offering groundbreaking solutions to address these challenges [7], and the paper proposes a blockchain-empowered privacy-preserving federated learning (PPFL) for a remote sensing image classification framework with the poisonous dishonest majority, which can defend against encrypted model poisoning attacks without compromising users’ privacy [29]. To enhance the execution speed and reduce the number of filters in low-cost embedded boards, the paper introduces Spatial Attention-based Filter Pruning (SAFP) for YOLOv4 and YOLOv7. This innovative approach leads to a significant improvement in processing speed without compromising the accuracy of real-time object detection [30].

These recent references reflect image classification’s dynamic and exciting directions in 2023. Additionally, image segmentation also plays a vital role in image processing and computer vision. The development of image segmentation methods is closely connected to several disciplines and fields, e.g., industrial inspection [31], intelligent medical technology [32], augmented reality [33], and autonomous vehicles [34]. Image segmentation is currently witnessing a remarkable surge in trends and techniques, with an expanding emphasis on leveraging deep learning models, such as graph neural networks (GNNs) [35], and attention mechanisms [36]. Furthermore, the combination of generative adversarial networks (GANs) [37] and reinforcement learning [38] holds promise for accomplishing more robust and adaptable segmentation models. As research continues to push the boundaries of image segmentation, the convergence of these cutting-edge techniques is poised to revolutionize not only the field of computer vision but also numerous industries reliant on accurate and efficient image analysis.

In the field of computer vision, the success of transformer networks in natural language processing tasks [39] has inspired researchers to apply this architecture to computer vision tasks, resulting in the development of Vision Transformers (ViT) [40]. In ViT, global attention is directed towards 16 × 16 patches of the entire image, focusing on the global salient features of the image, and resolving the long-range dependency among the image’s contents. This approach effectively balances the incorporation of the global context and maintains computational efficiency. The potential of vision transformers to address various problems in computer vision has grabbed the interest of numerous researchers, and their applications continue to be a subject of ongoing research. Vision transformers (ViTs) have introduced a novel and groundbreaking approach to object detection when integrated with frameworks like DETR (detection transformers (DETRs) [41]. ViTs capture intricate spatial relationships between objects within an image, and ViTs effectively analyze how different objects relate to each other and their context within the scene [42]. The utilization of DETR in conjunction with ViTs is of utmost importance in object detection because it enables a comprehensive understanding of the image, which is essential for identifying and localizing objects. This approach allows for simultaneous attention to all parts of the image, significantly improving the accuracy and efficiency of the detection tasks. The significant advances in object detection achieved using this approach highlight its potential. Researchers are actively exploring ways to refine these models for various computer-vision challenges, making object detection a promising and evolving field of research.

In the future, we anticipate a substantial change towards more advanced and multi-functional image processing techniques. This transformation will be propelled by rapid developments in artificial intelligence, machine learning, and computational capabilities. Key trends include the integration of deep learning models into image processing systems, resulting in increased accuracy and versatility in tasks such as image segmentation, classification, image super-resolution, and object detection. The increasing demand for real-time and efficient processing will give rise to novel solutions that leverage hardware acceleration and cloud computing. In summary, the future of image processing is likely to be a dynamic and transformative journey that will redefine how we interact with and harness visual data across various domains.