A Defect Recognition Method for Switchgear Infrared Images Based on Improved YOLOv9s

Currently, switchgear infrared inspection is primarily based on manual assessment of temperature overrun faults, exhibiting a limited degree of automation. This paper proposes the integration of a target recognition algorithm to identify equipment within the switchgear cabinet and the utilisation of the relative temperature difference method to assess equipment faults. In the domain of target detection, Yolov9s (You Only Look Once Version 9 Small) exhibits a notable advantage and can be employed for the recognition of infrared image defects. However, there are limitations to the detection accuracy and the model is too large to be effectively deployed on the mobile side. Accordingly, this paper proposes an enhanced lightweight YOLOv9s-mobilenetv3-adowncbam network to efficiently achieve lightweight implementation. The primary enhancements entail substituting the backbone with a lightweight MobileNetv3 (Mobile Neural Network Version 3) network architecture, incorporating a novel Adown downsampling and CBAM (Convolutional Block Attention Module) attention mechanism at the neck layer, and constructing a YOLOv9s-mobilenetv3-adown-cbam model. The experimental results demonstrate that the enhanced network model exhibits a $\mathbf{9 5. 9 \%}$ high-precision performance in the [email protected] indicator, while the GFLOPs value is only 25.8. In comparison to the original YOLOv9s model, the accuracy is enhanced by 5.4%, and the computational load is reduced by approximately 19.9%. This enables the model to operate effectively on mobile devices with enhanced performance.