Image Positioning Methods Of Catenary Components Based On Structure Prior Information And Model Compression

With the development of image processing and deep learning technology,the intelligent detection of catenary support suspension devices is also making continuous progress,but there are still many problems in the current detection technologies.1)When facing the components of catenary support suspension devices with complex structure and a large number,due to the scale difference and complex layout,the current detection methods are difficult to position small-size components,so most of them only detect certain components,and the positioning efficiency is not high.2)In railway field detection,due to the space limitation and the need for real-time detection,there are high requirements on the model size,memory occupation and running time used by detection methods,while the existing traditional detection models are difficult to be applied because they don’t meet the requirements.In this paper,the deep learning image processing technology is applied to the positioning detection of catenary support suspension components,and the image sample database of catenary support suspension components is established by collecting images from the highspeed railway detecting vehicle.Because of the above two problems,this paper puts forward two different solutions.1.Aiming at the difficulty of positioning small size components in the catenary,this paper proposes a new positioning model for 12 types of components of catenary support suspension device,which is called catenary support components detection network based on multi-level feature and structural prior information CSCSIN.The model is composed of three parts,including feature extraction using low and high-level features,feature inference using structural priori information between components,and regression and classification,this model achieves high precision positioning of 12 components with large scale differences.2.1 Aiming at the requirements of catenary detection model size and real-time detection,this paper adopts compression acceleration technology to solve these problems.1)Three different pruning methods,hard filter pruning,asymptotic soft filter pruning and channel sparse pruning,are selected to accelerate the model compression of traditional deep learning network model YOLOv5.On the premise of meeting the actual detection requirements,the parameter scale and memory footprint of the model are reduced,and the detection speed of the model is increased.2)Network distillation technology is applied to transfer the knowledge of a large network model to a small network model,to retain the advantages of the small size model and improve its positioning accuracy.2.2 According to the requirements of catenary detection model size and real-time detection,the traditional deep learning detection models have high positioning accuracy but large model size,while the classical lightweight models have fast detection speed and small model size but low accuracy.In this paper,the lightweight network is selected as the backbone network to combine with the traditional deep learning model YOLOv5,which makes up for the shortcomings of both and integrates the advantages of both,so as to reduce the complexity and memory access cost under the condition that the model has a certain accuracy.