ChWDTA: Channel-wise Wavelet-Domain Transformer Attention and Entropy Modeling for Learned Image Compression
Researchers have developed a new method for learned image compression called ChWDTA, which integrates channel-wise wavelet transforms into transformer attention mechanisms and entropy modeling. This approach sparsifies channel covariance for attention projections and improves rate-distortion performance. The ChWDTA scheme achieved significant BD-rate reductions across multiple test sets, demonstrating the benefits of incorporating wavelet transforms into hybrid CNN-transformer architectures for image compression. AI
IMPACT Introduces a novel technique for image compression, potentially improving efficiency and quality in multimedia applications.