Research On Mobile Phone Motherboard Appearance Defect Detection System Based On Deep Learning

With the rapid development of technology and chip industry,the iteration speed of smart phones is getting faster and faster,and China’s mobile phone market ranks among the top in the world.Therefore,the industry for the mobile phone motherboard PCB production process appearance defect detection demand is more and more urgent.At present,the detection of the appearance defects of PCB of mobile phone motherboard is mostly done by manual,but manual visual inspection is too dependent on human experience and mental state,and the integration of mobile phone motherboard is getting higher and smaller,so the visual inspection defects will have high rate of missed inspection,large omission and low efficiency.Therefore,in view of the demand for appearance inspection in the production of mobile phone motherboards,combined with the target detection network in deep learning,this paper designs and develops an appearance defect detection system of mobile phone motherboards based on deep learning instead of manual visual inspection.First of all,the characteristics and detection requirements of the appearance defect detection of mobile phone motherboard are studied,the characteristics of the defects produced in the production process of mobile phone motherboard are analyzed,and the difficulties of the current appearance defect detection of mobile phone motherboard are analyzed.The overall structure framework of the system scheme and the detection flow of the defect detection algorithm are designed.Secondly,in order to simplify the defect detection problem,the deep learning target detection network can be better applied.The image is preprocessed and a high resolution image is obtained by splicing image based on SURF algorithm.The location of the image was located by template matching based on shape,and the relationship between the location point and the detection region was established by special coordinate affine,so all detection regions were located.Image clipping is adopted to intercept the region to be detected from the original image as the input of the deep learning network.Then,the structure of R-CNN series network is studied,and Res Net101 is used as the backbone network based on Faster R-CNN network,so that it can extract more feature information.FPN network is introduced to improve the recognition ability of small target defects.In Faster R-CNN original network,using NMS for violence screening will cause the problem of defect detection,so soft-NMS algorithm with linear weighting is proposed to replace the original NMS algorithm.In order to better input the captured image of the region to be detected into the convolutional network,the rotation method is used to expand the data to avoid uneven positive and negative samples.The gray filling method is used to adjust the size of the original data to avoid image distortion and feature loss caused by the original data entering the convolutional neural network.The processed data were input into the improved Faster R-CNN network for model training and corresponding curves were drawn.The detection effect of the model is verified and the location of defects is determined by reference method.Finally,the software of the defect detection system was developed and applied to the established hardware platform for verification.The system was deployed on the actual production line.Through the production data collected for five consecutive days,the average processing time of the deep learning defect detection system was 8156 ms and the average pass-through rate was 95.194%,which met the production demand.Figure [61] Table [9] Reference [80]...