HRNet: Local-Spatial Feature Fusion Network for Texture Recognition

Texture recognition is a challenging and complex task within the realm of computer vision, because texture images contain local texture detail information and spatial distribution of feature. Some existing Convolutional Neural Network (CNN) approaches emphasize local feature distribution and encoding, relying solely on pooling operations to characterize global spatial feature. However, these approaches overlook the global spatial feature that cannot be fully captured by pooling operations. To address this problem, a Local-Spatial Feature Fusion Network named HRNet is proposed, HRNet consists of three prominent modules: histogram layer module, residual pooling module and feature fusion module. Histogram layer module constructs histograms to depict the feature distribution of local regions. Residual pooling module exploits residual connections and feature learning to extract spatial features. Finally, feature fusion module integrates the output features of the two modules to form a comprehensive texture representation. HRNet achieves accuracies of $75.1 \%, 87.2 \%$, and $85.4 \%$ on three challenging and complex texture recognition datasets DTD, FMD, and MINC-2500, respectively. The proposed HRNet not only demonstrates outstanding performance in capturing local and global texture features, but also achieves advanced performance on challenging datasets.