Few‐shot electrical equipment image recognition method based on an improved two‐stage fine‐tuning approach

Abstract In the process of electrical equipment detection based on deep learning, insufficient image samples of the target equipment will lead to detection failure. To solve this problem, an object detection network and two‐stage fine‐tuning approach based on You Only Look Once (YOLO)v4 is proposed in this paper to achieve image recognition of electrical equipment under the condition of small samples. Using the two‐stage and dual‐network method as the training strategy, the data‐rich base class samples are used to train the sample classifier based on the modified cosine similarity in the base class training stage, and the fine‐tuning is carried out in the small sample new class training stage. In the training part, the improved Retinanet network is used for coarse detection and the YOLOv4 network with Convolutional Block Attention Module (CBAM) attentional mechanism module is inserted for fine detection. The experimental results show that the average accuracy of the proposed method under the settings of 5‐shot, 10‐shot, and 30‐shot is 31.6%, 34.3%, and 52.8%, respectively, which greatly improves the ability of electrical equipment identification under the condition of few‐shot.