Frequency-Domain Guided Image Classification With Large Model Assistance

Image classification technology has made significant advancements, but methods tailored for specific image classification, such as sport image, remain inadequate. This is primarily constrained by two factors. First, the image quality of sport images varies greatly, and many rare sports have very few images. Second, most sport image classification techniques directly transfer image classification methods from other domains, with few approaches specifically designed based on the unique features of sport images. To address this problem, we devise a Frequency-domain Guided sport image classification method with Large model Assistance, named FGLA. Specifically, we design two main modules for FGLA. The first module combines Fourier Transform and Wavelet Transform to embed the frequency-domain information into the original image as additional channels, which converts the image into a three-dimensional cuboid, incorporating more comprehensive information. This allows the model to assign different weights to each channel, thereby enhancing image classification performance. The second module leverages the powerful image generation capabilities of large models to augment the dataset with more images, especially for rare sports. Additionally, it directly generates frequency-domain images to enhance the generalization of the classification model. We analyze the effectiveness of FGLA across many models and several classic sports datasets. The results indicate that FGLA achieves the highest accuracy in sport image classification. Moreover, this method also demonstrates strong generalization capabilities and can be adapted to other image classification tasks.