Hazard Detection for Power Transmission Channel Based on Texture Enhancement and Context Awareness

Hazard detection of power transmission channel generally obtains images of power transmission channels through surveillance cameras at fixed locations on the power transmission site. Due to the distance between the detected target and the surveillance camera, the target usually occupies a small proportion in the image, which poses a great challenge. This paper proposed a network model for hazard detection caused by construction machinery targets for power transmission channels, named TC-YOLO. This model is based on the YOLOX model and improved firstly by designing a texture enhancement module on the basis of feature fusion network, so that textural details from shallow feature maps can be used to compensate the losing details for small targets. Secondly, a depthwise context awareness module is added to the feature fusion network. With the help of context information around the target, it can assist the model in determining the category of small targets with similar appearance Finally, the boundary box loss function is adjusted by adding an exponential parameter to improve the accuracy of boundary box regression. Experimental results show that the proposed TC-YOLO model achieved higher detection accuracy for small construction machinery targets compared to other network models on experimental dataset. Furthermore, its detection speed can meet the demand for real-time detection.