RFSFeat: Advanced Feature Extraction and Recognition for Zhuang Brocade

To address the challenge of low feature extraction accuracy in complex scenes, particularly for Zhuangjin, we propose a novel shared feature extraction method named RFSFeat, which incorporates a feeling field attention mechanism. This innovative approach is designed to significantly enhance the accuracy of feature extraction for Zhuangjin. The feeling field attention mechanism is seamlessly integrated into a decoupled training process that follows a description-then-detection sequence. Central to this method is the newly designed feeling field attention convolution (RFAConv), which effectively addresses the issue of parameter sharing within convolutional kernels, thereby substantially boosting network performance. Furthermore, we introduce a shared coupling bridge that facilitates efficient information exchange between the description network and the detection network during training tasks. This mechanism indirectly achieves joint optimization, thereby enhancing the learning efficiency of local features. In our comprehensive feature extraction experiments, the RFSFeat method demonstrated remarkable performance, achieving a feature point extraction accuracy of 51%. When compared to the traditional PosFeat method, RFSFeat showed superior performance across various conditions: it improved matching accuracy by 0.14 percentage points under varying lighting conditions, by 0.13 percentage points under different viewpoint changes, and achieved an overall matching accuracy improvement of 0.13 percentage points. These experimental results underscore the significant advantages of the RFSFeat method and its excellent application potential in the feature point extraction of Zhuangjin, marking a notable advancement in the field.