There are many small targets in infrared images of substations with complex environment, resulting in low accuracy of existing detection algorithms. Therefore, this paper proposes an infrared image detection method for substation equipment based on improved Centernet. Firstly, taking Centernet as basic model, the FPN structure is introduced into the upsampling network to fully use the feature information of small targets, so as to solve the problem that small targets are difficult to be accurately detected; Then, in order to improve the detection robustness of the network in complex environment, an attention mechanism is embedded in the backbone network resnet50 to increase the attention of network to core targets; Finally, the training strategy of center point offset loss and width and height loss is replaced by CIOU loss to accelerate network convergence and improve training effect. The experimental results show that the method in this paper can have better detection effect in both small targets detection and complex environment detection, and the detection accuracy is improved by 3.1%, reaching 92.7%, which is more accurate than existing methods such as Faster R-CNN, and has certain reference value in infrared image detection of substation equipment.