Learned Image Compression in Low Bitrate based on Relevant Properties prioritized using Generative Adversarial Network

Abstract

The primary idea behind lossy image compression is to reduce image size. Image information is separated into three categories in image compression techniques: irrelevant, relevant, and redundant image data. Lossy image compression removes irrelevant and redundant information, leaving only relevant information. Existing image compression methods face challenges in maintaining visual quality at low bitrates below 0.1 bpp. Recent learned compression methods have shown promise in extremely low bitrate compression. However, most techniques treat the image holistically without considering the relevance of different visual properties. In this research, we developed an efficient learned compression approach using GAN architecture that prioritizes the compression of relevant features containing important visual information. Two important features - color and grayscale - were extracted from images and assigned varying priorities for compression. We choose two images features based on their importance in different areas to change lossy images, reduce image size, and generate image quality based on user needs. The proposed method utilized two networks in the generating module: compressor and decompressor. The compressor network takes a feature image and compresses it based on the priority set. The decompressed features were then combined to reconstruct the images. Experiments conducted on COCCO dataset showed the proposed technique with color prioritization achieved 30.5 PSNR and 0.84 SSIM at 0.02 bpp, significantly improving over baseline. The proposed Learned Image Compression in Low Bitrate based on Selecting important Feature to give Priority using Generative Adversarial Network was compared to GAN for Extreme Learned Image Compression and GAN for Fidelity-Controllable Extreme Image Compression. Our work results demonstrate the efficacy of the proposed technique in improving visual quality and compression ratio by selective feature prioritization, while avoiding commonly faced problems like artifacts at very low bitrates. Selective prioritization of relevant features like color, and grayscale enables efficient low bitrate image compression while improving visual quality. The approach can facilitate storage and transmission of images in bandwidth-limited applications.