Research On Mobile Phone Screen Surface Defects Detection Based On Deep Learning

In recent years,China has successfully surpassed Japan and South Korea to become the country with the largest comprehensive production capacity of Liquid Crystal Display(LCD)in the world.At present,the mainstream display panels are two types of thin film transistor array liquid crystal displays and organic light emitting diodes.High throughput requires a high-efficiency defect detection system.However,because the defects on the surface of the mobile phone screen are close to the background grayscale difference,the boundary lines of the defects are blurred,the proportion of the defect screen is small,the size and shape of the defects vary,imaging will have uneven brightness.The traditional machine vision defect detection method brings an insurmountable detection bottleneck.At the same time,traditional machine vision requires experienced engineers to manually design a fixed feature matching operator to match the extracted feature input.The fixed design operator is not robust in the face of increasingly rich types of defects and changing backgrounds.This thesis proposes a deep learning method applied to LCD screen defect detection.The main work of this thesis is as follows:1.Simulation sample generation(Virtual Sample Generation,VSG).Due to the imbalance between classes of datasets collected from industrial sites and the small number of samples.In order to achieve effective data set enhancement,the author studies the physical causes and imaging characteristics of defects,and then realizes the development of automatic defect generation and automatic labeling tools.VSG augments the defect data by a factor of 300 using random parameter changes.VSG also takes into account the uneven gray level of the area caused by lighting and environmental factors,and simulates the real shooting environment during the generation process.It solves the problem of imbalance between sample classes,small number of samples,labor-intensive manual labeling of massive data.2.Defect detection algorithm based on deep learning.(1)In order to meet the detection capability and detection speed of the industrial field at the same time,an LCD-YOLOv3 model that takes into account the detection accuracy and detection speed is proposed as the overall detection networks.There is no obvious grayscale difference compared to the surrounding background,so the mobile phone screen defects have low contrast.After the experimental verification,the multiscale feature map fusion is carried out on the original network and a layer of output prediction scale is added.The shallow high-pixel low-semantic information and Deep low-pixel and high-semantic features are fused to improve the detection ability of small targets.The k-means++clustering algorithm is used toobtain the anchorframe which is more consistent with the data distribution.Experiments show that the average accuracy(Average Precision,AP)of point flaws,line flaws,and group flaws has increased by 4.79%,2.27%,and 1.5%respectively compared with the previous ones.The detection ability of the fusion network for small defects has a certain improvement effect.(2)Due to the large difference in shape between defects,the detection effect is good for circular defects that occupy two-thirds of the data set,and the detection effect for defects such as thin line scratches is not good.The analysis is due to the fixed convolution shape.The ability to capture defects is not enough,so the Deformable Convolutional Networks(DCN)module is added to the residual block inside the feature extraction network.According to the experiment,the AP of the original line defect is 62.32%,which is improved based on the deformable convolution.The AP of the network LCD-DCN-YOLO line defect reaches 80.52%,which is 18.2%higher than that before the deformable convolution is added,and the mAP of the whole class is increased by 7.15%,indicating that the deformable convolution is enhanced.The ability to extract defects with large shape changes greatly improves the average correct rate of line defects.(3)Liquid crystal screen defect detection algorithm based on improved loss function.The original loss function is improved.When calculating the localization loss,the IOU loss cannot reflect the distance between the two boxes and the gradient direction is lost when the real box and the predicted box do not intersect or intersect but are completely included.,which cannot be optimized if this occurs.In order to consider the overlapping area between the real frame and the predicted frame and the distance between the two centers at the same time,and to weaken the shortcoming that the IOU is not sensitive to distance,the loss function based on the distance intersection ratio loss is introduced to optimize,by adding the predicted frame and the real frame The distance penalty term of the center point of the box to take into account the intersection area between the real box and the predicted box and the distance between the two centers.The final effect is a 1.87%increase in overall mAP.To sum up,the LCD screen defect detection network proposed in this thesis has improved the detection ability for various defects,and has certain reference significance for actual industrial field detection.