Research On Local Feature Enhancement Methods Of Catenary Detection Images

The normal operation of the catenary ensures the normal power supply of the high-speed railway.For the fault detection of the catenary,the non-contact detection technology of the catenary support device components developed in recent years has greatly improved the fault detection accuracy and reduced the workload of manual monitoring.The accuracy of noncontact detection technology depends on the quality of the catenary image.Affected by hardware and environment,the catenary images obtained by the collection vehicle have quality problems such as low brightness and contrast,which will reduce the accuracy of the noncontact detection technology and bring a burden to manual monitoring.In response to the above problems,this paper seeks a deep learning enhancement method suitable for catenary images to achieve the enhancement of catenary images,improve the detection accuracy of non-contact detection technology and reduce the interference to manual verification.Firstly,based on the catenary images collected from the Beijing-Guangzhou line and in accordance with the PASCAL VOC data set format,a data set containing 11 types of key components such as Messenger wire base,insulator and Rotary double ear is established.The data set established is adopted to train the Faster RCNN-based target detection network,and a stable model is obtained,and the location of the components of the catenary support device is achieved.This model is used to evaluate whether the accuracy of non-contact detection technology is improved after the catenary image is enhanced.Secondly,in order to reduce the interference to manual verification caused by low-quality catenary images and improve the detection accuracy of non-contact detection technology,this paper adopts the enhancement method based on Generative Adversarial Network to realize the enhancement of catenary images.This method regards the enhancement of brightness and contrast as a migration task.The mapping between the source domain images(this paper refers to images with low brightness and low contrast)and the target domain images(this paper refers to images with normal brightness and contrast)is learned,and the enhancement is realized indirectly.Considering the actual working situation that the collection vehicle cannot obtain the paired data set,this paper constructs an unpaired data set for network training(The images in the source domain and the images in the target domain are not one-to-one correspondence),and the enhanced model named model E is obtained.The Faster RCNN trained by enhanced images can achieve better location effect for catenary support components,and the detection accuracy of non-contact detection technology has been improved.Then,to improve the situation that some pixel information is lost in the enhancement process of the Generative Adversarial Network,this paper adopts the enhancement method based on deep reinforcement learning to realize the enhancement of catenary images.Specifically,the model E is first used to enhance 1100 catenary images,and the enhanced images and the original images form a paired data set.Based on this paired data set,a deep reinforcement learning method is used to change the pixel value of the low-brightness catenary image to enhance the catenary image directly.Finally,considering the particularity of catenary image,and the actions executable by the agent are redesigned to achieve better enhancement effects.The experimental results show that directly changing the image pixels to achieve enhancement can effectively avoid the loss of pixel information,and the improved enhancement action achieves a better enhancement effect than the original network.