The work aims to improve the detection efficiency and accuracy of wire mesh. An improved network EfficientNetV2 for wire mesh surface defect recognition was proposed. Firstly, the backbone structure of the network was changed, and operations such as channel splitting and channel conversion were introduced before and after the feature extraction module to increase network capacity and improve feature utilization. Secondly, the classifier of the network was redesigned, and the extracted high-level semantic information was compressed layer by layer to reduce the feature loss and improve the classification accuracy. Finally, an image acquisition system was built to construct a wire mesh defect data set. According to the experimental results, the accuracy, precision and specificity of the improved network model on the data set were 99.43%, 99.42% and 99.88% respectively, and the image recognition time was 27.5 ms, which enhanced the defect recognition effect. The method has high accuracy and good practicability in wire mesh defect detection, and can also provide reference for defect detection of other similar products.