Efficient Lossy Image Compression and Reconstruction With DCT-Optimized Residual Compression Framework

Image compression reduces file size while preserving visual quality, improving storage and transmission efficiency. Traditional lossy methods struggle at low bitrates, introducing artifacts that degrade image quality. While deep learning offers potential solutions, its adoption is hindered by challenges like training instability, sensitivity to hyper parameters, and mode collapse. To address these challenges, this research proposes a novel DCT-Optimized Residual Compression Framework (DCT-RCF) that integrates frequency domain transformations with an enhanced deep learning architecture. The framework leverages the Frequency-Enhanced residual Autoencoder (FERA-Net) which combines residual learning and an autoencoder with Discrete Cosine Transform (DCT) to efficiently separate essential visual structures from redundant spatial information. The proposed approach enhances compression efficiency by encoding high impact frequency components while discarding insignificant details, ensuring high quality reconstructions with minimal artifacts. The reconstruction process utilizes Inverse Discrete Cosine Transform (IDCT) and transposed convolution to restore spatial domain details while minimizing reconstruction errors. The framework is evaluates using Kodak and CLIC datasets with performance metrics including Mean squared Error, Peak Signal to Noise Ratio, Compression Ratio and Multi-Scale Structural Similarity Index Measure. Experimental results demonstrate that the proposed DCT-RCF achieves a compression ratio of 44.26:1 significantly outperforming the traditional and DL based compression methods by achieving PSNR of 42.38dB and MS-SSIM of 0.999 ensuring superior image reconstruction quality even at low bitrates making it highly effective solution for advanced image compression applications.