Image Compression May Improve Classification Accuracy

Figure 9 compares ResNet-18, ResNet-50, and MobileNet-v2 under different JPEG compression levels. Identity and PNG serve as baselines.

Across all three models, a small amount of compression improved accuracy at first. Once compression became stronger, accuracy dropped. The likely useful effect is not compression by itself, but the removal of some high-frequency variation before too much task-relevant information is lost.

The source code and experiment results are available on GitHub.