Research And Application Of Part Defect Detection Method Based On Semantic Segmentation

As the basic unit of industrial products,the quality of parts directly determines the overall quality of industrial products.However,due to factors such as equipment failure and harsh working conditions during the production and processing process,defects on the surface of the parts often occurs.The surface defects on the parts not only affect the beauty of the parts,but also have a negative impact on the service life of the product,thus resulting huge economic losses to the factories and users.At present,the surface defect detection of domestic parts is mainly based on manual detection,while others rely on machine vision detection of feature engineering.However,manual inspection has low efficiency,strong subjectivity,and unstable inspection quality;machine vision inspection methods rely too much on artificial design features and are poor in flexibility.This paper takes the sealing ring as an example to study the surface defect detection of parts,and conducts generalization experiments on other parts data sets.The main work content is as follows:(1)The construction of the seal ring image data set.According to the defect characteristics and structural characteristics of the sealing ring,we build the sealing ring image acquisition platform.Then,we utilize the image acquisition platform to collect data on the sealing ring samples,which are further cut,annotated and enhanced to construct the sealing ring data set.(2)We propose a defect detection method based on improved U-Net network.Taking the U-Net network as backbone,we improve the network according to the characteristics of the seal ring image,thus improving the ability of the network on segmenting the seal ring defect area.Our proposed detection algorithm has the following improvements:1)By injecting image features of different scales into each convolution module on the encoder,the network is able to obtain the multi-scale information of the input image.Meanwhile,we add a dilated convolution layer to increase the receptive field of the network and to reduce the loss of detailed image information in the down-sampling process.Besides,we introduce an attention mechanism to achieve high-precision segmentation of the defect area;2)We perform feature fusion of the decoder output to achieve the complementarity of multi-scale feature information,which is conducive to fusion of semantic information and location information and improves the model’s segmentation accuracy for smaller defects.Besides,we use other parts data sets to verify the generalization of our proposed algorithm.(3)We propose a defect detection method based on recursive U-Net network.Inspired by the second observation and thinking mechanisms of the human visual system and the recursive feature pyramid,we propose a defect detection method based on the recursive U-Net network.Our proposed method enables the recursive U-Net to obtain the feature representation of rich feature information,thus improving the segmentation accuracy of the region with defects.Then,we evaluate the generalization of the recursive U-Net network on other part data sets.