Interpretable Object Detection For Remote Sensing Power Based On Feature Enhancement And Screening Mechanism

In recent years,the development of power vision technology based on digital image processing has been rapid,and research on power target recognition and fault detection is an important application of power vision technology.In the context of demand for fault detection and safe operation and maintenance,the stable operation of power equipment in power plants is one of the key links to ensure the safety of the power grid.Existing power target detection mainly uses infrared and thermal imaging technology to monitor and evaluate the safety performance of small power equipment in real time.However,large power equipment such as cooling towers and units have high volume and complex space,which makes it difficult to observe their health status quickly and recognize and detect them using traditional image technology.In order to improve the fault detection capability of large power equipment,the use of remote sensing image analysis and detection technology,which is different from digital image processing technology,is a feasible means.This dissertation focuses on the task of remote sensing power target detection using deep learning algorithms,First,preprocess the collected images,then perform object detection on the enhanced images to obtain images of large-scale power equipment.Finally,use class activation mapping to visualize features and provide assistance for downstream fault analysis tasks.The main research content and achievements are as follows:Firstly,A preprocessing fusion algorithm based on homomorphic filtering and Retinex image enhancement is proposed.The satellite-acquired images are subjected to image enhancement and defogging processing,which optimizes the interference caused by clouds and fog on the image quality and improves the clarity of remote sensing image features.The experiments show that this algorithm can effectively improve the quality of image features and enhance texture details.Secondly,A HRTFS-YOLOv5 object detection algorithm based on a filtering mechanism is proposed.To address the problem of low detection accuracy in power target detection,the network model is improved.Firstly,the CBAM attention module is added to the backbone structure to enhance the network’s ability to detect small objects.Secondly,a feature selection mechanism is designed to eliminate redundant information generated in remote sensing images and reduce the computational cost of network training.Thirdly,to speed up the network training speed,an optimization loss function based on logarithmic gradient acceleration convergence is designed,and the Euclidean distance between the predicted center coordinate and the true target center coordinate is added as a cost function to the regression loss function.Experiments on a remote sensing power dataset show that the improvements with different strategies have significantly improved the accuracy of object detection.Lastly,A class activation mapping(CAM)based Mask-Grad-CAM interpretable model is proposed.To address the lack of interpretability of deep learning-based artificial intelligence methods,especially the determination of detection decision points,feature visualization is performed.The model first uses backpropagation to update the weights and overlays them with the original image to obtain a feature heatmap.On this basis,a mask branch is added,and the gradient weights generated in the first step are used to calculate the weighted sum as a mask.The final heat map visual key feature image is obtained by subtracting it from the original image and iteratively iterating.The algorithm explains the classification criteria and decision points of the network model.Experimental results show that the average detection rate of the class activation mapping with the mask strategy is over 75%,achieving good results in multi-object detection.