Abstract: In recent years, image data augmentation methods based on multi-image mixing have achieved promising results. However, approaches that fully adopt patched pixel regions from source images may generate low-quality synthetic samples in certain scenarios, requiring further improvement. To address these issues, we propose DynamicMix, a dynamic pixel-level mixing algorithm. A local pixel mixing strategy was introduced to select appropriate cropping regions from source images while preserving partial pixel values, achieving localized pixel-level blending. To mitigate the impact of varying cropping sizes on synthetic samples, a dynamic pixel-level mixing mechanism was proposed by associating the cropped image patches with adaptive mixing ratios, which ensured that the preserved pixel ratio from the source image dynamically adjusted according to the size of the cropped region. This approach prevented scenarios where excessive loss of critical features from source images-due to large cropping areas-lead to significant discrepancies between label assignments and actual content. Experiments on four datasets demonstrate that the proposed data augmentation method enhances both classification performance and model robustness. Notably, when applied to CIFAR-100 and Mini-ImageNet datasets with ResNet-34, the method achieves Top-1 accuracy improvements of 1.00 and 1.14 percentage points, respectively, compared with the CutMix baseline.