Surface defect detection network of power equipment based on DCGAN

The application of computer vision and deep learning algorithms in the field of industrial defect detection is getting deeper and deeper. Aiming at the problems of complex background interference and difficulty of micro-scale defect feature extraction in the detection of surface defects of electric power equipment, the research proposes a kind of deep convolutional generative adversarial network that integrates the spectral normalization and coordinate attention mechanism. The algorithm solves the problem of gradient disappearance and pattern collapse in the multimodal data training of traditional generative adversarial network by dynamically constraining the spectral paradigm of the weight matrix, while embedding the coordinate attention module in the high-resolution stage of the generator, which directionally enhances the recognition accuracy of geometric structural features for tiny defects on the surface of power equipment by decomposing the spatial coordinate information. Experiments show that the proposed network achieves 93.2% precision and 94.3% recall on Coovally dataset, and the detection speed reaches 46.7FPS, which significantly improves the accuracy and real-time performance compared with the comparison model, and provides an efficient algorithmic solution for power equipment inspection.