Application Research Of Automobile Wiring Harness Appearance Detection Based On Object Detection

Recently,CNN(Convolutional Neural Networks)have gradually replaced traditional handcrafted features and played an important role in the field of computer vision.CNN process images using alternating convolutional layers and pooling layers.Compared with traditional methods,it can achieve higher accuracy in image classification challenges.Benefiting from the ability of CNN to learn feature representations from images,the field of object detection has developed rapidly,and different types of detection frameworks have emerged.The detection accuracy of the target detection method is gradually improving,so it is widely used in various fields,such as face detection and pedestrian detection.In the industrial field,although technologies such as machine vision can assist in the identification of defects,the main body of defect detection is still manual detection,and inspectors directly observe or use specific tools to inspect the product.The manual detection method is limited by the working speed of human beings.What's more,as the working time increases,the status of the detection personnel decreases,then inspectors may miss some defective products.The using of manual inspection consumes a lot of human resources,while the detection accuracy is not high enough.Thus,the method of defect detection needs to be upgraded.Aimed at the disadvantages of the traditional defect detection and machine vision methods,this thesis uses Mask R-CNN to perform overall appearance detection of the wiring harness,combines the object detection with the wiring harness appearance detec-tion organically,and achieves application innovation.Aiming at this goal,the thesis does image acquisition,cleaning and labeling on wiring harness samples.Finally,the thesis constructs a high-quality wiring harness appearance detection dataset.In the wiring harness dataset,there are some problems which influence the appear-ance detection.Thus,this thesis proposes a multi-strategy Mask R-CNN model.The fea-ture pyramid networks is used to detect instances of different sizes in the harness dataset.The deformable convolutional layer is applied to learn the complex appearance of some instances.What's more,the thesis combines the GA(Guided Anchoring)method with Mask R-CNN for the problem of large changes in the size and aspect ratio of the sample instances.After combining different methods,the multi-strategy model has a good detec-tion accuracy for the wiring harness appearance detection task,and can effectively detect the sample instances in the image samples.At the same time,in view of the situation that Mask R-CNN's mask head is difficult to perform high-resolution instance segmentation tasks,this thesis combines the semantic segmentation method to improve the mask head,fuses features of different levels together,and proposes a innovative dual input mask head.The dual input mask head greatly im-proves the instance segmentation accuracy of Mask R-CNN for large-size samples in the wiring harness dataset.