Deep Learning-Based Approaches for Enhanced Image Compression in Computer Vision Applications

Avinash A.; Quevedo Lezama C.R.; Murugesan R.; Anitha K.; Bharathi C.R.; Abid Y. · International Conference on Distributed Systems, Computer Networks and Cybersecurity, ICDSCNC 2024 (2024) · DOI: 10.1109/icdscnc62492.2024.10939783

Resumen

Image compression is required for many computer vision applications, but traditional methods like JPEG and PNG have trouble preserving image quality at high compression levels. The study suggests a deep learning-based method for efficiently compressing images while maintaining visual information, utilizing convolutional neural networks (CNNs). Preparing the dataset, choosing CNN architectures (such as autoencoders and GANs), training with loss functions like MSE, and assessing using PSNR and SSIM metrics are all part of the process. According to the results, the deep learning approach outperforms JPEG (18:1) and PNG (6:1) in terms of compression ratios for complicated textures, reaching up to 25:1. The deep learning approach's PSNR values are 35.2 dB, while its SSIM scores are 0.92 in contrast to JPEG's (0.87) and PNG's (0.84) values. The process preserves visual quality while offering increased storage and transmission efficiency.

Instituciones: Universidad Nacional Mayor de San Marcos