Optimized Deep Learning-Based Image Compression Using Convolutional Autoencoders

Abstract

This study presents an optimized image compression framework based on deep learning, specifically a convolutional autoencoder. Traditional compression methods such as JPEG and PNG rely on fixed mathematical transformations, which often lead to quality degradation at high compression ratios. To address these limitations, the proposed system adopts a data-driven approach that learns compact and efficient image representations through end-to-end training. The framework integrates convolutional encoding, latent space quantization, entropy coding, and decoding mechanisms to ensure efficient storage and transmission. Experimental evaluation was conducted using benchmark datasets, and performance was assessed using metrics such as Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and compression ratio. The results demonstrate that the model achieves a compression ratio exceeding 133:1 with minimal perceptual quality loss, maintaining stable PSNR and SSIM values across varying compression strategies. Furthermore, the model exhibits fast convergence, strong generalization capability, and real-time processing performance, making it suitable for deployment in resource-constrained environments. Despite challenges such as computational training cost and minor loss of high-frequency details, the proposed approach significantly improves compression efficiency and visual quality. These findings highlight the potential of deep learning–based methods in advancing next-generation image compression systems.