Research On Image Recognition Algorithm For Bolt Loss Fault Of Freight Car

In recent years,the railway transportation has been developed rapidly in China and the traditional manual inspection cannot meet the growing requirement of the freight car security system.The trouble of moving freight car detection system(TFDS)has improved the quality and efficiency of the security inspection greatly.It can identify the trouble states of the key parts' images of freight car automatically,which has an important practical value and theoretical significance.At present,the image recognition algorithms in the freight car trouble detection area are rare and most are based on traditional methods.In this paper,an effective and reliable automatic image recognition algorithm for the bolt loss fault of freight car is established.The key technologies of the preprocessing,feature extraction and classifier steps in the implement process are studied.The main research contents and results involve the following aspects:Firstly,the improved ranked-order based median filter(IRAMF)algorithm is proposed to improve the noise reduction performance in the image preprocessing step.The multi-level noise detection strategy,including the whole detection,the local detection,the neighborhood similarity detection and the edge detection methods,is designed to identify the noise points correctly.The adaptive filter strategy,including the local texture direction synthesis filter of the sub windows,the Manhattan distance weighted mean filter and the minimum size window with the nearest signal points,is designed to improve the filter capability.It can adaptively adjust the size,shape and output calculation method of the current filter window according to the noise density and texture structure characteristic.Experimental results on typical field images show that it can effectively protect the texture details of the original image with strong noise filtering capability.Secondly,the completed direction local binary patterns(CDLBP)operator and the label weight ranked dominant LBP(LWR-DLBP)method are established to solve the image feature extraction problem.The new operator includes not only the sign and magnitude change trend of the local difference but also the comparison results of the mean gray and magnitude value of the neighborhood and the whole image.Considering the occurring probability,the distribution variance of different patterns and the class information,the new mode selection method is designed to select the most discriminative patterns.Taking the train bolt images for experiments,the performances of CDLBP operator and LWR-DLBP method are analyzed.Experimental results indicate that these improvements are very effective.Thirdly,to solve the parameter optimization problems in the engineering practice,the improved teaching-learning-based optimization(ITLBO)algorithm is proposed.The optimum strategy of group selectionis designed to replace the greedy strategy of individual selection.The adaptive teacher phase is designed with the adaptive learning step,the adaptive learning difference and the Lévy mutation of teacher.The adaptive learner phase is designed with the self-learning,the cooperative learning and the exchange learning.These improvments are designed to accelerate the converge speed,improve the searching ability and skip the local optimum.Compared ITLBO with other optimization algorithms on the fixed and variable dimensions benchmark functions,experimental results prove its superiority with faster convergence speed and better search results.Finally,the design and application of the images recognition system of the freight car trouble are studied.The structure and main technical specification of TFDS are declared.The implementation scheme based on Gabor transform are established.The response results of the original image in different scales and directions are acquired and the CDLBP operator is used to extract the features of each channel independently.The performances of feature vectors extracted from different channels under LWR-DLBP method are analyzed.Different weights of the predicted labels of different channels are optimized by the ITLBO algorithm.Finally,the performance of the proposed recognition algorithm is tested on the center plate blot lost images,the coupler yoke bolt lost images and the safety chain dropped images.Experimental results indicate that it can identify the trouble state of the key parts' images of freight car effectively.