Enhancing Image-Based JPEG Compression: ML-Driven Quantization via DCT Feature Clustering

JPEG compression is a widely used technique for reducing the file size of digital images, but it often compromises visual quality. The purpose of this research is to explore a novel approach that combines machine learning, discrete cosine transform (DCT) feature clustering, and genetic algorithms to customize image compression methods. The goal is to enhance visual quality while maintaining an appropriate bit-rate. In this study, an auto-encoder neural network is utilized to extract DCT features from images. These features are then clustered, and optimized quantization tables for each cluster center are generated using a genetic algorithm. The resulting tables are assigned to their respective clusters, enabling the preservation of visual quality during compression. Experimental evaluations were conducted on 1800 random images using this machine learning-based approach. The results demonstrate superior visual quality compared to traditional JPEG compression, while maintaining comparable bit-rates. The research shows significant improvements in peak signal-to-noise ratio (PSNR) by 2.34 dB and structural similarity index (SSIM) by 1.26%, indicating enhanced image quality. The findings of this research highlight the potential of combining machine learning, DCT feature clustering, and genetic algorithms to customize image compression techniques. The proposed approach enables effective image compression with improved visual quality preservation and maintained bit-rates. This research contributes to the advancement of image-based methods in achieving optimized image compression.