Research On Toy Defect Detection Based On Image Segmentation And Matching

With the promotion and application of intelligent manufacturing in industrial production,industrial inspection is facing a transition from manual visual inspection to machine vision.The detection algorithm is the core part of machine vision inspection,which is directly related to the stability,efficiency and speed of machine vision inspection.Therefore,it is of great significance to study stable and efficient detection algorithms.With the support of the key project subject of the Changsha Science and Technology Bureau,this paper studies the defect detection of a certain set of toys.According to the characteristics of multiple varieties and small batches of this set of toys,from two aspects of traditional image processing methods and deep learning methods,an appropriate defect detection algorithm is proposed to completed the defect detection of toys.The main research work of this article is as follows:Based on the traditional digital image processing method,the color image is distributed in the HSV(Hue-Haturation-Value)space,combined with the Otsu algorithm to accurately segment the region of interest in the toy image.On this basis,through SIFT(Scale-Invariant Feature Transform)algorithm extracts the feature points of the image,and combines the gray information to match the regions.A toy image defect detection model based on HSV spatial threshold segmentation and SIFT algorithm matching is constructed,which can quickly and accurately segment toy images for matching,and the model has a high accuracy rate..In view of the poor generalization ability and flexibility of the model obtained by traditional image processing methods when the data distribution changes,a defect detection model based on deep learning was developed.In this model,U-Net is used to construct an image segmentation network,and CNN(Convolutional Neural Networks)is used to construct an image matching network.In addition,in view of the inaccurate U-Net segmentation edge and the occasional category prediction error,IUnet is proposed in terms of network structure and loss function to improve it,and the IUnet-CNN model is obtained.The research results show that the IUnet-CNN model can accurately segment and match toy images,and is superior to the defect detection model based on traditional image processing methods in terms of detection accuracy and generalization ability.Introduce transfer learning to the IUnet-CNN model.When a new data set and a small data volume detection task appears,use model-based migration to fine-tune the trained IUnet-CNN model on the new data set to obtain a new Detection model.The research results show that the new model obtained by using the original IUnet-CNN model for transfer learning is more accurate than the new model obtained by retraining the new data set.