Identification And Recovery Of Catenary Image Of Abnormal Exposure Based On Deep Learning

With the vigorous development of rail transit business,traditional catenary detection methods have been unable to meet today’s needs.Therefore,video-based intelligent detection methods for catenary system fault detection have become a development trend.At present,most of the research on abnormal state detection of catenary are concentrated on the algorithm level.By improving the detection accuracy of the algorithm,the safety guarantee capability of the railway is enhanced.However,when the vehicle camera’s perception of the ambient light intensity reaches saturation or the camera adopts the wrong exposure operation,the phenomenon of abnormal exposure of the image will occur,and it is difficult for the anomaly detection algorithm to accurately identify such images.Therefore,this paper mainly starts from the inspection image level to identify and restore abnormally exposed images.For abnormally exposed images,this paper optimizes the image quality evaluation model to achieve the scoring of abnormally exposed vehicle catenary images.And for the local hanging string area in the image,the random forest algorithm is used to realize the identification of abnormal exposure of the local hanging string image of the catenary.Combined with the fault detection algorithm of the hanging string,the effectiveness of the method in this paper is verified.The image inpainting method based on the context encoder is improved,and the restoration model of the overexposed hanging string is constructed to realize the restoration of the overexposed hanging string image.At the same time,the practical application effect of the method in this paper is verified by the fault classification and detection of the restored hanging string.The main research contents of this paper are as follows:(1)Due to the low quality of abnormally exposed images,this paper proposes a method of image quality evaluation to realize the identification of abnormal exposure of vehicle catenary images.In this paper,the existing image quality evaluation methods are studied.Since it is difficult to confirm the actual situation corresponding to the abnormally exposed images,the method of non-reference image quality evaluation is introduced to evaluate the quality of abnormally exposed images.According to the local abnormal exposure phenomenon in the image and the characteristics of many redundant image features,an attention module,a local feature extraction module and a lightweight module are introduced into the convolutional neural network to complete the optimization of the quality evaluation model.In comparative experiments on public image quality evaluation datasets,the accuracy of our model outperforms other models.The scoring test is performed on the vehicle catenary image dataset to verify the effectiveness of the model.(2)Aiming at the problem of identifying abnormal exposure of local dropper images of catenary,this paper proposes an identification method for abnormal exposure of droppers based on random forest.First,the exposure value of the image and the maximum value of the minimum gray level of the row are calculated,and then the classification threshold is set to determine whether there is an abnormally exposed dropper in the image.Since the method of manually setting the classification threshold is not universal,this paper uses the exposure value of the image,the minimum gray value of the row,and the minimum gray value of the row as the input features of the random forest model to realize the selection of the classification threshold.And completed the identification of abnormal exposure of dropper images.It is verified that the random forest model has good classification ability through the comparative experiment,and combined with the fault detection algorithm of the dropper,the fault detection experiment of the dropper is carried out,and the effectiveness of the method in this paper is verified.(3)Aiming at the restoration problem of overexposed dropper images,an overexposed dropper image restoration model based on context encoder is proposed.A new mask image is designed according to the characteristics of overexposed dropper images.In order to make the generated images more realistic,the encoder-decoder structure in the model is improved by combining the skip connection idea in the U-Net network.In order to better discriminate the restored images with irregular missing regions,the discriminator structure of the model is improved by combining the idea of subregional discrimination in the Patch GAN network.It is verified that the model has good image recovery ability through comparative experiments,and the fault classification test is carried out on the recovered dropper images,which proves the effectiveness of the recovery model.