Real-time Image Detection Algorithm For Key Components Of Train In Complex Background

With the rapid development of computer technology and machine vision theory,fault image detection method for key parts of train plays a more and more important role in the field of railway.On-line dynamic detection system of TFDS(Trouble of moving freight car detection system)which is popularized by the Ministry of railways is gradually replacing traditional manual inspection,and it will become more automated,intelligent and unmanned.Based on the dynamic fault detection system(TFDS)of train operation,this paper combines the background structured model,confidence map,correlation filtering and sparse combination learning.etc to study the intelligent real-time image detection of the lock plate deflection fault,the chains loss fault of hand brakes and the defect of the plate bolt.The main contents and related achievements are as follows:1.Researching on the anomaly detection of the lock plate deflection based on the background structured model.Considering the abnormal condition of the lock plate deflection under complex background,the central transformation feature descriptor was proposed to construct the structured background model,which can effectively reduce redundant image information and form the sparse representation of image data.In addition,considering the complex and changeable light and the strong interference of texture and noise from surrounding area,a hierarchical detection framework was built based on the principle of structural significance,and real-time reliable location of the lock plate was realized by combining the reliability atlas and the correlation filtering technology.On this basis,the two-dimensional attitude discrimination,the image edges were extracted by regional segmentation,and the edge interference was removed by gradient weighting.Finally,the accurate detection of the lock plate deflection state was realized by using radon transform.2.Researching on the chains loss fault detection of hand brakes based on sparse combination learning.In view of the complete dictionary representation in the traditional sparse representation,it is found that the appropriate combination of atomic base from redundant space will cause waste of time,which makes the detection speed of actual application requirements can’t been satisfied.Based on the structured characteristics of hand brake chains,the structural model was constructed and the pattern features of each part were represented by the temporal and spatial gradient in this paper,and the better sparse representation of the image data was realized by principal component analysis method.The sparse combination learning based on limited base vector combinations was used instead of the traditional dictionary search to reconstruct the test data,which has greatly improved the detection speed.At the same time,considering the lag of the base vector combination of learning to the current data reconstruction process,the reliability of the detection system was improved by using the combined set online real-time update strategy.3.Researching on the fault detection of plate bolt deletion based on the abnormal detection theory.Considering the problem of low reliability of the traditional projection transformation combining with template matching method and a lot of time was taken when using the neural network during the detection process of the plate bolt.The background structured model was constructed by central transformation according to the structure characteristics of the plate bolt images.Considering the interference of complex and changeable background noise,according to the hierarchical detection framework,the real-time location of the heart disk region was realized on the basis of the theory of confidence atlas and correlation filtering.Further,the bolt was accurately positioned according to the progressive region segmentation method,and the real-time reliable detection of the plate bolt was realized by sparse combination learning.