Recent Advances in Image Compression Techniques: From Traditional Methods to Deep Learning Models

Abstract

Image compression is a fundamental and critical component of modern multimedia and communication systems, enabling efficient storage, processing, and transmission of visual data across various platforms such as mobile applications, cloud storage, streaming services, and remote sensing systems. The rapid growth of high-resolution imaging technologies has increased the demand for advanced compression techniques that can reduce data size while preserving perceptual image quality.
Traditional image compression techniques such as JPEG, JPEG2000, and various transform-based coding methods have been widely used for decades. These approaches rely on mathematical transformations including Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), quantization, and entropy coding to reduce redundancy in image data. Although these methods are computationally efficient and well-established, they often face limitations in achieving an optimal balance between compression ratio and visual quality, particularly in complex images with rich textures and fine details. Additionally, artifacts such as blocking and blurring are commonly observed at high compression levels.
In recent years, deep learning-based image compression techniques have emerged as a powerful alternative to traditional methods. Approaches based on Convolutional Neural Networks (CNNs), autoencoders, variational autoencoders (VAEs), and transformer architectures have demonstrated significant improvements in compression efficiency and reconstructed image quality. These data-driven models learn compact latent representations of images and optimize compression directly based on reconstruction loss and perceptual metrics, leading to superior performance compared to handcrafted transform-based methods.