About this Research Topic
Image/video quality assessment, enhancement, and compression are fundamental topics in the low-level computer visions which have witnessed rapid progress in the last two decades. Due to various degradations in the image and video capturing, transmission, and storage, image and video might incur a series of undesirable effects, such as low resolution, low light condition, rain streak, blackness, raindrop occlusions, and high-frequency detail loss, etc. The estimation and recovery of these degradations are highly ill-posed. With the wealth of statistic-based frameworks, i.e. traditional Maximum-a-Posteriori (MAP) Estimation and Rate-distortion joint Optimization (RDO), and learning-based tools, e.g. deep networks, meta-learning, and adversarial learning, many recent deep-learning-based methods have shown their significant performance gains over traditional non-deep methods.
Existing image/video quality assessment, restoration and compression methods have remaining issues to be addressed. The challenges arise from the following aspects:
1) As enhancement/compression models are trained on the training data collected from limited scenes or occasionally synthetically generated ones, their performances might sharply degrade on real-world images and videos when there are domain gaps between real applications and training data.
2) Existing losses used for the model training are proven to be misaligned with the human vision experiences, more efforts are expected to design better measures to describe human vision experiences closely.
3) Existing methods are mainly designed for human vision. With the big data captured from smart cities and the Internet of Things, more applications expect to feed the data into machines. It would be the new critical issue to build new approaches to enhance and compress images/videos for both humans and machines.
4) Existing models include more than millions of parameters, which pose obstacles to real applications.
Topics of interest include (but are not limited to):
• Novel architectures, models, and approaches for image and video quality assessment, restoration and compression.
• Novel theories, optimization methods, training skills for training models and networks for low-level vision.
• Computationally efficient networks for image/video quality assessment, restoration and compression.
• Learned enhancement and compression models for humans and machines.
• Deep learning-based techniques that improve the performance of existing codecs and standards.
• Quality assessment methods that are well aligned to human visual perception.
• New enhancement/compression methods guided by perceptual measures or analysis tasks.
• Explainable deep learning for image/video quality assessment, restoration and compression.
• Unsupervised/semi-supervised learning methods that learn to enhance/compress images/videos with fewer labels.
• Robust methods trained with domain adaptation or elaborately designed constraint to learn from noisy labels collected from real-world data.
• Compression for compact descriptors, deep features, semantic features.
• Collaborative or Adversarial Learning for Machine Vision.
• Scalable and Distributed Architectures for Machine Vision.
Existing image/video quality assessment, restoration and compression methods have remaining issues to be addressed. The challenges arise from the following aspects:
1) As enhancement/compression models are trained on the training data collected from limited scenes or occasionally synthetically generated ones, their performances might sharply degrade on real-world images and videos when there are domain gaps between real applications and training data.
2) Existing losses used for the model training are proven to be misaligned with the human vision experiences, more efforts are expected to design better measures to describe human vision experiences closely.
3) Existing methods are mainly designed for human vision. With the big data captured from smart cities and the Internet of Things, more applications expect to feed the data into machines. It would be the new critical issue to build new approaches to enhance and compress images/videos for both humans and machines.
4) Existing models include more than millions of parameters, which pose obstacles to real applications.
Topics of interest include (but are not limited to):
• Novel architectures, models, and approaches for image and video quality assessment, restoration and compression.
• Novel theories, optimization methods, training skills for training models and networks for low-level vision.
• Computationally efficient networks for image/video quality assessment, restoration and compression.
• Learned enhancement and compression models for humans and machines.
• Deep learning-based techniques that improve the performance of existing codecs and standards.
• Quality assessment methods that are well aligned to human visual perception.
• New enhancement/compression methods guided by perceptual measures or analysis tasks.
• Explainable deep learning for image/video quality assessment, restoration and compression.
• Unsupervised/semi-supervised learning methods that learn to enhance/compress images/videos with fewer labels.
• Robust methods trained with domain adaptation or elaborately designed constraint to learn from noisy labels collected from real-world data.
• Compression for compact descriptors, deep features, semantic features.
• Collaborative or Adversarial Learning for Machine Vision.
• Scalable and Distributed Architectures for Machine Vision.
Keywords: Deep-Learning, Image/Video Enhancement, Image/Video Compression, Image/Video Quality Assessment, Video Coding for Machines
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
