Research On Infrared Image Intelligent Detection Algorithm Of Power Equipment Based On Deep Learnin

Long-term operation of power equipment in complex environment may lead to equipment failure caused by equipment body temperature rise.Abnormal heating phenomenon is one of the classic manifestations of equipment failure in operation.It is necessary and important to realize the detection and treatment of abnormal heating state of power equipment for maintaining the safe operation of the system,discovering and responding to potential accidents.Infrared image detection is an effective method to detect abnormal heat generation of power equipment.This paper discusses the intelligent detection of infrared images of power equipment from three aspects: image data expansion,lowillumination image enhancement and object detection.(1)In the aspect of image data expansion,an infrared image combination expansion method based on supervised single sample power equipment is proposed,which combines and expands the key contents.In the single-sample expansion,the traditional single-sample expansion method will cause a large amount of redundant information and similar features in the expanded image,which will lead to the problem that the training has been closed.This paper conducts comparative experiments and analysis based on the five methods of flipping,cropping,scaling,contrast and noise in the supervised single-sample expansion method.Experiments show that the effect of the double combination expansion method is better than that of any single expansion method,and the contrast + flipping and contrast + adding noise are the best combination expansion methods among the three combination expansion methods.(2)In terms of low-light image enhancement,a low-light infrared image enhancement method for power equipment based on improved Retinex-Net is proposed,including the improvement of Retinex-Net and superpixel segmentation and reconstruction.1)In terms of network improvement,aiming at the problem that the original infrared image has low brightness and the target details are unclear,the traditional enhancement method cannot adaptively adjust different brightness regions.In this paper,by improving the network smoothing loss function and adding adaptive brightness correction and component fusion module,a preliminary enhancement of low-light infrared images of power equipment based on improved Retinex-Net is proposed.2)In terms of superpixel segmentation and reconstruction,aiming at the problem that the contrast between foreground and background is not obvious after preliminary enhancement,this paper introduces the method of superpixel segmentation and reconstruction.Firstly,the superpixel segmentation method of Maximal Similarity based Region Merging(MSRM)and Super-pixels Extracted via Energy-Driven Sampling(SEEDS)is used to extract the target of interest in the preliminary enhanced infrared image,and then the multi-scale infrared image fusion method of hybrid guided filtering is used to reconstruct the extracted target image to obtain the final enhanced image.The experimental results show that the proposed enhancement method can improve the brightness of the image without causing obvious color offset and distortion,retain the image details more clearly,improve the comprehensive quality of the image,and help to improve the accuracy of the target detection model.(3)In terms of object detection,an improved YOLO xs power equipment infrared image object detection method based on transfer learning is proposed,and the improvement of detection network is the key content.Aiming at the problem that traditional YOLO xs networks have small models but low detection accuracy,this paper proposes corresponding improvement strategies from three aspects: transfer learning,feature extraction and loss function.Experiments show that transfer learning can effectively improve the learning speed of the model.Adding the Adaptively Spatial Feature Fusion module to the tail of the feature fusion layer can filter useless information while retaining useful information and integrating it,improving the multi-scale feature fusion ability.The confidence loss and localization loss functions of the network can be improved to further improve the detection accuracy and speed of the model.