Vision-Language Feature Fusion and Wavelet Attention for Fine-Grained Visual Recognition

Abstract

Fine-grained visual recognition requires models to distinguish subtle differences among visually similar categories. Recent vision-language models have shown strong semantic representation ability, while wavelet-based convolutional networks improve sensitivity to frequency-domain details. This article proposes a hybrid feature fusion framework that combines semantic-guided visual encoding with multi-scale wavelet attention. The framework first extracts visual features through convolutional and transformer-based encoders, then enhances texture, contour, and lesion-related information through wavelet transform convolution. A semantic guidance module introduces category-level descriptions to improve discriminative learning and interpretation. Finally, an attention module refines channel and spatial features before classification or caption generation. The method is designed for fine-grained recognition tasks such as remote sensing scene analysis, melanoma image recognition, breast cancer DCE-MRI segmentation support, medical imaging interpretation, and industrial defect identification. The article highlights the complementary relationship between semantic representation and frequency-aware visual detail. By integrating these two feature sources, the framework aims to improve both recognition accuracy and interpretability in complex visual domains.