Multi-scale frequency domain learning for texture classification

Recently, methods for modeling images in frequency domain have attracted widespread attention. Frequency methods transform image into spectrum as model input by defining a set of basis functions, where each point in frequency domain represents image’s projection at the corresponding frequency, which helps to understand structure, texture, and other basic features of the image. Studies have shown that frequency learning methods based on fixed small-scale Discrete Cosine Transform (DCT) basis functions outperform spatial domain methods. However, frequency learning also faces the challenge of feature saturation bottleneck, particularly fixed DCT basis functions makes the frequency learning model more prone to reaching saturation. Therefore, we propose a multi-scale frequency domain learning approach that utilizes multi-scale DCT basis functions to extract diversified frequency features. Our method employs sampling and cropping techniques to extend the frequency features from a single-scale DCT basis functions mapping to a multi-scale mapping. This ensures that the image projection exists at as many frequencies as possible so that advanced features can be extracted over multiple frequency ranges to overcome the saturation bottleneck of network models. Experimental results demonstrate the superior performance of our approach in texture classification tasks.