Core Concepts
Improving image compression efficiency through channel-wise feature decorrelation.
Abstract
The paper discusses the shift from traditional codec modules to Deep Neural Networks (DNN) for image compression. It introduces a method to enhance compression efficiency by fully utilizing existing DNN capacity through channel-wise feature decorrelation. Three strategies are proposed and evaluated to optimize transformation and context networks, resulting in up to an 8.06% improvement in compression efficiency without added complexity. Experimental results on established LC methods validate the effectiveness of the proposed approach.
Stats
Experimental results show an improvement in compression with a BD-Rate of up to 8.06%.
Training time increases by 5.32% to 9.97% compared to the baseline.
Quotes
"The proposed method consistently improves compression performance without adding complexity."
"Experimental results confirm significant gains in compression efficiency."
"The proposed approach enhances image reconstruction without increasing computational complexity."