Identification Method of Top Plate Pallet of Underground Roadways in Coal Mines Based on Lightweight Faster R-CNN

Abstract Convolutional neural networks are widely used in target recognition with their excellent feature extraction capabilities. The Faster R-CNN algorithm for object detection using region suggestion networks has high detection accuracy, but the drawbacks of large number of parameters and high computational effort hinder its deployment on resource-constrained mobile and embedded devices. In order to realize accurate and fast recognition of pallets on the top plate of the underground roadway in coal mines, a pallet recognition method based on lightweight Faster R-CNN is proposed, considering the requirements of real-time and implantability of the model. ResNet18 is selected as the feature extraction network, and the depth separable convolution is used instead of ordinary convolution to operate on the input image, which reduces the number of parameters and computation amount of the network, and thus improves the speed and efficiency of the network feature extraction. On this basis, the model accuracy is improved by the knowledge distillation method, in the training process, the ResNet50-Faster R-CNN is used as the teacher network, the lightweighted ResNet18-Faster R-CNN is used as the student network, and the training of the student network is supervised by using the soft labels output from the teacher network in order to improve the accuracy of the student network, and the performance of the student model obtained by the method is close to that of the teacher model, but the number of parameters and the cost of computation are both smaller than that of the teacher model. Based on the above proposed improvement method, tests are carried out in a simulated alleyway to verify the effectiveness of the method, and the proposed algorithm is compared with three classical algorithms. The experimental results show that the proposed improved algorithm has better performance in pallet recognition, compared with the original Faster R-CNN algorithm, the lightweight improved algorithm model occupies 59.47% less space, and the recognition time of a single picture is only 5.54% of the original algorithm, which reaches 22.7ms, and the mAP value is only reduced by 3.58%. It can be seen that the improved algorithm basically maintains the original recognition accuracy on the basis of the model memory and algorithm operation is greatly reduced, the detection speed has been improved, can effectively meet the real-time requirements of the pallet image recognition, and helps the pallet recognition model deployed on mobile and embedded devices.