AYOLOv3-Tiny: An improved convolutional neural network architecture for real-time defect detection of PAD light guide plates

Abstract

Light guide plates (LGPs) are the main component of the backlight unit of liquid crystal display (LCD) devices, and defective LGPs directly affect the display effect of LCDs. In view of the features of portable Android device (PAD) LGP images, such as complex texture background, low contrast, different defect sizes and various defect types and their optical properties, light-spot distribution, defect formation principle and imaging characteristics, this paper proposes an AYOLOv3-Tiny network for the defect detection of LGPs. First, by combining overlapping pooling and the spatial attention mechanism, the overlapping pooling spatial attention module (OSM) is constructed to replace the traditional convolution operation of the YOLOv3-Tiny backbone network. Overlapping pooling can improve the accuracy and prevent overfitting, and the spatial attention mechanism can help the network better extract defect features. Second, a dilated convolution module (DCM) is constructed in the detection branch. The module can expand the receptive field of the convolution kernel and improve the detection ability of large defects by integrating the dilated convolution into the residual network structure. Third, a large number of experiments based on the self-built dataset PAD LGP SDD are carried out. The experimental results show that the mean average precision (mAP) and F1-score of the LGP defect detection system can reach 99.50% and 99.61% respectively, and the detection speed can reach 144 fps. Finally, by testing on the PAD LGP images with defects, it is verified that the proposed network meets the application requirements of high precision and real-time online detection of PAD LGP images.