Visual Saliency Detection Method And Its Engineering Application In Rail Transit

Visual saliency detection refers to human attempts to construct systems that simulate human vision and cognition for obtaining regions of interest(saliency regions)in the scene.Researchers use visual saliency detection to highlight salient regions in a scene,which can be used as a preprocessing technique for numerous visual tasks and their applications.With the rapid development of artificial intelligence technology,applying machine learning methods(including deep learning technology)to visual saliency detection and its application in defect detections has become a current research hotspot.This dissertation focuses on the saliency detection of RGB and RGBD images,and introduces machine learning and deep learning algorithms to improve the image scene saliency ability of traditional RGB image saliency detection models,as well as the detection performance of RGBD image saliency detection models in complex scenes.In addition,this dissertation also applies saliency detection method to the engineering defect detection of key components in rail transit,focusing on the detection difficulties of vehicle bogie axle box cover bolts and track line fasteners,a core defect detection method with visual saliency detection is constructed.The specific research content is indicated as followed:(1)To address the issue of traditional RGB image saliency detection models not being able to effectively highlight prominent targets in image scenes,an RGB image saliency detection model based on iterative bootstrap learning ensemble model is constructed,which mainly includes two stages:saliency optimization and saliency integration.The main points can be indicated as:a saliency regression and prior saliency map will be constructed through the random forest algorithm and the existing saliency detection model,the fusion of the initial saliency map and the rough saliency map will be achieved by iterative bootstrap learning to obtain final saliency map.The experimental results on three public datasets show that the proposed model has achieved better performance and can be used to improve the performance of existing models.(2)In response to the serious performance degradation of RGBD saliency detection models in the processing of complex scene images,an RGBD image saliency detection model based on attention guided feature ensemble network is constructed.The key points can be expounded as:enhanced deep features are fused and generated by deploying the attention module onto RGB features and Depth features.Meanwhile,the edge information of RGB and Depth branches are fused to generate high-quality significance maps.According to the experimental results on five public RGBD datasets,the proposed model has achieved excellent performance on all datasets.(3)A bolt looseness detection method based on saliency detection of identification lines is proposed to address the engineering bottleneck of detection technology for bolt status on rail vehicles.This solves the problem of bolt feature recognition and achieves effective recognition of the bolt status of rail vehicle axle box covers.The main points can be indicated that the deep semantic information of the bolt identification line image is extracted by fusing the saliency detection model PFA and the attention model SE.Moreover,the shallow edge features of the bolt identification line image are extracted by using the dual attention model DA,so as to build the A~2-PFN identification line saliency detection model to extract the bolt identification line features and generate the saliency map.Based on this map,the bolt status can be identified by the angle table method.The accuracy and effectiveness of the proposed method in detecting the bolt status of vehicle axle box covers are experimentally demonstrated.(4)In response to the engineering bottleneck of detection technology for the status of track line fasteners,a sample generation based on track fastener detection method is proposed,which solves the problem of imbalanced samples in the detection model and achieves effective detection of track line fastener status.The key points are mentioned as:a multi-stage information fusion based on fastener significance detection network(MIF-FSDNet)is proposed,which specifically constructed by Conv Ne Xt-T encoder.Moreover,a decoder is constructed to obtain accurate prospects for fasteners by a residual attention module RX-Att,CA and ECA attention models.According to the sample generative model and Res Net network,a fastener detection model with strong stability and robustness is obtained,and the accuracy of the fastener detection model is verified through experiments.This dissertation constructs two saliency detection models for RGB and RGBD images,which effectively improve the performance of visual saliency detection.At the same time,visual saliency detection has been successfully applied to the detection of rail vehicle bolts and line fasteners.The research work in this dissertation is expected to promote the development of visual saliency detection technology and its application in the field of rail transit detection.