Comparative Analysis of Image Encoders and Compression Effects on Machine Task Performance

In recent years, there has been a sharp increase in automated analysis of images by machines. It is expected that the volume of images consumed by machines will greatly exceed the volume of images consumed by humans. The utilities of traditional image coding algorithms for machine consumption are not sufficiently explored. To address this issue, we conducted a study of compression efficiency for machine consumption of JPEG, JPEG XL, WebP, AVIF, HEVC, and VVC codecs. Test results show that the use of classical image codecs such as JPEG and WebP could lead to unsatisfactory performance of convolutional neural networks in object detection accuracy, especially at lower bitrates. State-of-the-art codecs VVC, JPEG XL, and AVIF could be more suitable for those use cases, despite the increased computational complexity.