Power Equipment Image Recognition Method Based on Feature Extraction and Deep Learning

Traditional image recognition methods for power equipment face challenges such as difficulty in distinguishing target features from background features and insufficient feature extraction capabilities. This paper proposes an improved attention mechanism-based network for image detection and recognition of power equipment. The proposed method introduces a target feature prediction strategy tailored to power equipment: it incorporates a learning mechanism for depth variation to extract deep semantic information from images; enhances the global structure learning network module by stacking convolutional kernels and removing pooling layers in the front-end network, thereby acquiring prior information rich in detailed and correlated image features of power equipment. Furthermore, a long short-term memory (LSTM) gate mechanism is employed to predict power equipment target features at different levels of image feature information, constructing an attention mechanism network based on the LSTM gating mechanism. Additionally, the method introduces a deep-shallow feature interaction strategy: it integrates shallow and deep feature information through matrix outer product operations, enabling the model to fully learn multi-level features of power equipment. Compared with traditional power equipment image recognition methods, the proposed approach enhances the recognition and extraction of detailed target features, accurately distinguishes blurred boundaries between background and targets, and improves the interaction between deep and shallow features, effectively increasing recognition accuracy in complex background environments. Experimental results show that, on image datasets of five types of power equipment—insulators, transformers, circuit breakers, transmission poles, and transmission towers—the proposed model achieves a recognition accuracy of 92%, which is 1.6% higher than that of the CvT model. Future research will focus on further enhancing the model’s robustness and generalization ability in complex scenarios. We plan to introduce a lightweight convolutional structure combined with a graph neural network mechanism to strengthen global context modeling and device structural awareness. This will enable efficient and interpretable identification and localization of power equipment in application scenarios such as automated substation inspections and real-time monitoring with drones.