Substation smoking behavior detection based on improved decoupling head

Abstract

In the complex environment of the substation, small objects such as long-distance cigarettes, cigarette boxes, and lighters have few imaging pixels and lack texture information, making it difficult for convolutional neural networks to extract small object features. In the case of multiple targets, the missed detection rate and false detection rate of small targets are high, and the fusion of model features is insufficient, making it difficult to accurately identify and detect. Aiming at the above problems, a multi-scale small target detection algorithm is proposed. In the prediction part of the network, a more effective decoupling head is designed. In addition, shallow features are introduced to improve the feature pyramid, extract small target features, increase the correlation between multiple targets, and prevent the loss of small target feature information. At the same time, a multi-layer attention mechanism is embedded in the backbone network to increase the regional features of invisible small targets and reduce the missed detection rate. In the post-processing stage, the Focal Loss loss function is introduced to increase the model's learning of positive sample targets and further reduce the rate of missed detection and false detection. The experimental results show that the method achieves a 𝑚𝐴𝑃@. 5: .95 of 0.6350 on the homemade smoking dataset, and 𝑚𝐴𝑃@0.5 achieves 0.9569. For the self-made multi-scene and multi-scale smoking data set, this model has advantages in detection accuracy compared with the current excellent target detection models such as YOLOX and YOLOv5. The experimental results show that the model method can realize the identification and detection of small targets such as cigarettes and lighters under multi-scale targets, which has a certain reference value for anti-smoking measures.