Automatic Inspection Method of Power System Based on Image Recognition

This paper constructs a model of power system and its defect detection network based on dense connection mechanism. In the feature extraction network, DenseNet-121 was optimized, Stem structure was added, and
$\mathbf{3}\times \mathbf{3}$
small convolution kernel and dual channel feature extraction method were used to construct a dense connection feature extraction network with reduced spatial dimension. Aiming at the end-to-end recognition problem of power equipment, the Dense connected network-Stem object detection model is put forward in this paper as an end-to-end recognition model for power equipment. The image of power equipment is preprocessed by normalization and tagging to establish the model training data set. The experiment compares which deep feature extraction network is used for automatic feature extraction, uses regional suggestion network network instead of traditional candidate box generation method, and integrates the whole process of box regression and classification into one model to achieve end-to-end training of the whole network model. The proposed method can not only meet the requirements of high-precision identification of power equipment, but also complete the end-to-end automatic learning and training. Compared with traditional target detection algorithms, the detection performance is greatly improved.