Report on Current Developments in the Research Area of Image and Video Compression and Processing
General Direction of the Field
The recent advancements in the field of image and video compression and processing are marked by a shift towards more adaptive, data-driven, and efficient methodologies. Researchers are increasingly focusing on integrating deep learning techniques with traditional compression algorithms to enhance performance and address specific challenges such as domain adaptation, frequency-based learning, and the handling of large-scale data like light field images.
Adaptive and Data-Driven Approaches: There is a notable trend towards developing adaptive methods that can dynamically adjust to different datasets and scenarios. This includes techniques for few-shot domain adaptation in learned image compression, which aims to improve the generalization capabilities of pre-trained models. Additionally, adaptive selection methods in Fourier compressed sensing are being explored to optimize both sampling and reconstruction processes, addressing the limitations of traditional optimization-based approaches.
Frequency-Based Learning: The integration of frequency-based learning into self-supervised and contrastive learning frameworks is gaining traction. These methods leverage frequency responses to enhance the pre-training process, leading to more robust models that can adapt better to downstream tasks. This approach is particularly promising for tasks involving natural images and time-series data.
Enhanced Compression Efficiency: Researchers are also focusing on improving the efficiency of compression algorithms, particularly for large-scale data like light field images. Novel techniques such as disentangled representation learning and asymmetrical strip convolution are being developed to capture complex spatial relationships and reduce bit rates significantly.
Fair and Comprehensive Evaluation: There is a growing emphasis on fair and comprehensive evaluation methodologies. This includes advocating for per-video basis BD-rate calculations in video codecs to ensure a more accurate comparison of performance. Additionally, the development of new test datasets like USTC-TD is providing a more diverse and representative benchmark for evaluating image and video coding schemes.
Noteworthy Papers
On the Computation of BD-Rate over a Set of Videos for Fair Assessment of Performance of Learned Video Codecs: This paper highlights the importance of fair evaluation metrics in video codecs, advocating for per-video BD-rate calculations to avoid misleading conclusions.
LFIC-DRASC: Deep Light Field Image Compression Using Disentangled Representation and Asymmetrical Strip Convolution: This work introduces innovative techniques for light field image compression, achieving significant bit rate reductions and enhancing the representation of intricate spatial relationships.
These developments collectively push the boundaries of image and video compression and processing, offering more adaptive, efficient, and fair methodologies for future research and applications.
