Deep Learning Based Lightweight Traffic Sign Detection Algorithm

With the development of deep learning,object detection,as a key problem in computer vision,has also been developed rapidly.The detection algorithm based on convolutional neural network makes use of its massive parameters and multi-layer nonlinear mapping,combined with data sets for training,and finally makes the model achieve better detection results.However,due to the large number of parameters,the model has high requirements on hardware and needs to be built on a high-performance server,while the effective hardware resources on the vehicle equipment cannot meet its application needs.Therefore,in view of the shortcomings of the general object detection model applied in such scenarios,The thesis conducted an in-depth study on the lightweight of the detection model and proposed an improvement plan.The specific work is as follows:(1)Based on baseline network YOLOv5,the model is improved according to the characteristics of small target size and high real-time requirement of traffic signs.Firstly,in order to improve the model’s attention to the important information in the feature graph,The thesis designs an attention grouping feature fusion network combined with the permutation attention mechanism.Then,Ghost convolution is introduced to replace the original convolutional operations in the baseline network,and lightweight improvements are made to the model.The attention packet feature fusion network is integrated into the Ghost Bottleneck CSP module,which replaces the C3 module in the baseline network.Finally,in order to avoid the loss of fine-grained feature information in the pooling process,the Maxpool of the baseline network is replaced with Softpool to retain more fine-grained features in the feature graph.By designing experiments and analyzing the results,the number of parameters of the model is reduced by 40%,and the accuracy is increased by 1.2%.(2)In view of the unbalanced distribution of traffic sign TT100 K data sets,the data sets are equalized by means of data expansion and data enhancement.Finally,16,582 images of traffic signs were obtained,including more than 42,000 traffic sign targets.Experiments were conducted with baseline network YOLOv5 to verify the effectiveness of data set equalization.The m AP of 45 types of traffic signs increased from 85.2% to87.7%.(3)Traffic sign detection algorithm based on improved knowledge distillation.In The thesis,the traditional knowledge distillation network is improved to make it suitable for traffic sign detection.Based on the knowledge distillation model,The thesis designs a local feature enhancement module to enable the teacher network to help the student network extract more refined features,so as to further improve the accuracy of the model.The thesis A cross-entropy based distillation classification loss function based on aggregation feature knowledge is constructed,and the regression loss function is combined to train the network.By designing experiments and analyzing the results,the accuracy of the model is improved from 88.9% to 92.6%,which proves the effectiveness of knowledge distillation.Figure 41 Table 7 Reference 74...