Traffic Sign Detection And Recognition In Natural Scenes

For the past decades,the autonomous driving technology has received great attention for its huge application prospects.However,providing algorithms to detection and recognition of the traffic signs more accurate and rapid is still one of the challenges for engineering in the 21 st Century.Herein,we hope to enhance the performance of the algorithms from the perspective of signal detection and recognition.In terms of traffic sign detection,this thesis uses a method based on Mask R-CNN.Using RestNet101-GFPN instead of the original feature extraction network,GFPN can enhance the original features of the image through the global fusion of semantic features,so that each feature layer in the pyramid can obtain the same information from other layers.In the process of model improvement,in order to make the model have a higher recall rate,adjust the RPN network to obtain high-quality regional suggestions,and use Soft-NMS instead of NMS to improve the detection accuracy of the model.Experimental results revealed that the improved Mask R-CNN model is significantly better than other models in terms of accuracy,recall rate,missed detection rate,etc.,and has a good traffic sign detection effect.In terms of traffic sign recognition,this thesis uses an Alex-Net-based recognition method.In order to make the improved I-AlexNet network model have better recognition accuracy and solve the gradient dispersion problem,2 convolutions are added based on AlexNet network model Layer and 2 Inception parallel convolutional layer structure;to prevent the model from overfitting,a Dropout layer is introduced after the second fully connected layer;to reduce the loss function of the model,the BGD algorithm is used to train the learning parameters of the network structure.Experiments on the GTSRB dataset proved that the parameters of the I-AlexNet network model consume less memory and have higher recognition accuracy.Furthermore,it also effectively shorten the training iteration time of the network model.Taken together,this method exhibited a good recognition effect.