Light Guide Plate (LGP) marker detection is a crucial procedure in LGP manufacturing quality control but a mass of bubbling, polluted and marker-free cases may arise out of detection in traditional image algorithms. Because the mass of bubble, polluted and unmarked lines makes people difficult to design features. Compared with other classification neural networks, DenseNet Convolution Neural Network (CNN) has fewer parameters and stable gradient convergence. Because DenseNet CNN uses the idea of feature fusion, the accuracy of image classification is guaranteed. Through the transform learning method, the weights of the trained DenseNet network are transferred to the bilinear-CNN algorithm for training, which improves the local attention of the convolutional neural network and improves the accuracy of image classification. The implementation results show that the proposed method is feasible. Compared with the V2-ResNet-101, the accuracy of proposed approach is increased to 95.53%, while parameter number is decreased by 97.2%, and average single image detection time drops by 25% in the proposed network structure.