Research Of Low Contrast Surface Defect Inspection Based On Convolution Neural Network

The product quality detection technology based on machine vision has been widely used in the field of intelligent manufacturing,and has achieved good results.However,the traditional image processing algorithms aren't good for low-contrast surface defect detection during factory automated production.It is one of the difficulties in the field of machine vision surface defect detection how to improve the performance of low-contrast defect detection algorithms.In this paper,the low contrast surface defects of magnetite are studied,and the defect detection algorithm based on convolution neural network is mainly studied,which is applied to practical production.It is the basic premise of all image algorithm research acquiring high-quality magnet defect images.The paper first selects the camera,lens,light source and sensor according to the characteristics of the magnet itself.In order to obtain high-quality images that can show the defects of various magnets,this paper attempts to use a variety of light source combination schemes on the basis of hardware selection,and then constructs the most ideal lighting scheme hardware structure diagram.In order to solve the problem of lack of samples of magnetic defects,this paper proposes a duck-filling algorithm,which is combined with the Mixup image augmentation method to increase the magnet sample data.The number of magnet images is enlarged form 800 effective samples to 4800.Next,this paper designs a defect detection algorithm based on a shallow backbone Convolutional Neural Network(CNN)for segmenting magnet defect areas in order to obtain more defect structures and gray-scale features.Experimental results show that this chapter algorithm's m Io U is 0.852 and detection accuracy rate is 92.08%.It has similar detection performance but faster detection speed than the detection algorithms based on deep backbone networks.In order to further improve the detection performance of small defects in the detection of magnetic defects,the paper proposes a magnetic defect detection algorithm based on the generative countermeasure network.This method uses a shallow backbone CNN as a generator network.The generator network is used to extract the structure and gray-scale features of learning magnet defects.In the discriminator network,a layer-by-layer jump connection structure is used to transfer multi-scale features,thereby improving defect detection performance.Experiments show that the magnet defect detection algorithm based on the generative confrontation network greatly improves the detection performance of small defects.The result of m Io U is 0.875,and the accuracy rate is 96.54%.In this paper,a defect segmentation network is designed to detect low-contrast magnet defects,and a generative adversarial network is used to improve the ability to detect small defects.The test results on the magnet data set show that the algorithm in this paper can meet the requirements of magnet defect detection.Therefore,the algorithm in this paper has certain theoretical and engineering application value in the field of defect detection.