Residual guided and cross-level feature interaction network for substation anomaly detection

Recently, YOLO-based object detection methods have achieved remarkable results in substation anomaly detection. However, existing detection methods lack target information with different scales of receptive fields during Backbone feature extraction, hindering the effective discrimination between targets and background amid complex background interference. Moreover, as network depth increases, the transmission of features from the backbone to the detection head is often subjected to information dilution, resulting in the inadequate utilization of shallow details and ultimately diminishing detection accuracy. To address these issues, we propose a Residual Guided and Cross-level Feature Interaction Network (RGCIN). Specifically, We propose a Residual Guided Feature Enhancement (RGFE) module that selectively amplifies multi-scale receptive field features by leveraging residual information, thereby augmenting the network’s capacity to accurately distinguish foreground objects. Furthermore, a Cross-level Feature Interaction Fusion (CFIF) module is designed to effectively integrate shallow texture features with deep semantic information through correlation-based queries among hierarchical features, thereby improving detection performance. Experimental results on the substation anomaly image dataset demonstrate that the proposed algorithm outperforms 14 state-of-the-art object detection methods. The code is released available at: https://github.com/wzd-l/RGCIN.