Improving long-tailed classification with PixDyMix: a localized pixel-level mixing method

With the continuous expansion of dataset size, the issue of long-tailed distribution has become increasingly prominent. Traditional approaches often favor head categories while neglecting the importance of tail categories. To address this limitation, this paper innovatively proposes the PDMLT (pixel-level dynamic mixing for long-tailed classification) method, the core of which lies in a pixel-level dynamic mixing image data augmentation technique called PixDyMix (pixel-level dynamic mixing). This technique intelligently adjusts mixing weights based on image cropping area, effectively preventing excessive loss of key pixel information during large-area cropping and improving the quality and label matching of newly generated samples. By generating higher-quality tail category sample images, it effectively increases the number of high-quality tail category samples, thereby enhancing the overall generalization ability of the model. Additionally, to overcome the limitations of existing resampling strategies in category weight allocation, we introduce an adaptive weight function to optimize the sampling process. This function can adaptively adjust the sampling weights of each category based on the degree of imbalance in the dataset, significantly improving the classification accuracy and stability of the model. Through comprehensive experimental validation on three standard long-tailed distribution datasets, our method demonstrates clear advantages and effectiveness.