Application Research Of Deep Learning In Defect Detection Of Mobile Phone Data Interface

Defect detection based on deep learning is a more popular detection technology in recent years.It has faster detection speed and higher detection accuracy.We improve the Faster RCNN network model,and propose a detection method based on the improved model,which can effectively improve the detection accuracy of small defects.And it can be applied to detect defects in the mobile data interface.The results show that the proposed detection method can significantly improve the detection performance.The main research work of the subject is as follows:1.An improved model is formed by combining the ResNet50 feature extraction network with the Faster R-CNN model.Due to the large amount of VGG feature extraction network parameters in the Faster R-CNN model,the detection speed of the network is slow.ResNet50 uses a residual network structure to reduce the amount of parameters and improve the detection speed.Experiments show that the improved Faster R-CNN model improves the detection speed.2.Incorporating the Feature PyramidNetwork(FPN)structure into the improved Faster RCNN model.Specifically,FPN is fused with the feature extraction module ResNet50 of the improved Faster R-CNN.Because small defects may be ignored when using ResNet50 for feature extraction,the feature map can be performed on different convolutional layers of ResNet50 by using the FPN structure.Extraction will retain small defects to a large extent,thereby improving the detection accuracy of the model.The results show that using the ResNet50 fusion FPN structure in Faster R-CNN can alleviate the problem of important information loss in the feature extraction process,and improve detection accuracy.3.Combining the RoIAlign network layer with the Faster R-CNN model to improve positioning accuracy.Because there are two quantization operations in RoIPooling in the Faster R-CNN model,the target positioning is biased.RoIAlign uses bilinear interpolation to overcome the shortcomings of RoIPooling.The results show that the detection accuracy is improved.Based on the above improved model,the mobile phone data interface defect experimental test was carried out.The test results show that using the proposed improved algorithm,the mean average precision(m AP)of two types of defects reached 91.17%,and the detection speed reaches 4.76 FPS.Compared with only using VGG as the feature extraction network,the m AP is increased by 24.72%,and the detection speed is increased by 3.09 FPS,and it is 2.58%higher than the m AP when only using ResNet50 as the feature extraction network,the detection speed has increased by 1.17 FPS.