Traffic Sign Recognition Algorithm Base On Compressed Convolutional Neural Networks

Traffic sign recognition is one of the important functions of intelligent transportation system,and it is also widely used in driverless cars,intelligent robots and other fields.Compared with general target recognition,the recognition of traffic signs in real scenes are easily affected by the illumination,occlusion,weather,shooting angle and distance.Traditional traffic sign recognition methods adopt hand-craft features,combined with some classifiers to identify the traffic signs.However,the expressive ability and robustness of the hand-craft features are not enough.The traffic sign recognition algorithm based on deep convolutional neural network can extract features from training data by an unsupervised method.But the automotive system can hardly meet the requirements of large convolutional neural networks for computing resources and storage space.The paper analyzes the advantages and disadvantages of some popular deep network compression algorithms,and studies the basic theoretical knowledge of convolutional neural networks.The paper improves the existing classic convolutional neural network model and compresses their parameters.Furthermore,the paper proposes a new deep network,which could deploy on resource-constrained devices for traffic sign recognition.Firstly,classical convolutional neural network models AlexNet and VGG16 have too many parameters and are difficult to transplant to mobile devices with limited running resources.Therefore,the model is pruned based on Taylor expansion to delete redundant feature map channels,then the parameters of modified model is trained with ternary quantized.The experimental results of channel pruned,ternary quantized parameter and combined compression for models are compared respectively.The experimental results on GTSRB show that the parameters of model is greatly reduced by combined compression,and high recognition accuracy is achieved on the GTSRB dataset.Secondly,the paper proposes two novel lightweight networks to obtain higher recognition precision while preserving less trainable parameters in the models by knowledge distillation.The first deeper network model is used as the teacher model,and the second shallow model is used as the student model.The teacher model is mainly composed of four new modules,which can get a good recognition rate on the CIFAR-10 dataset.The new module uses some 1×1 and 3×3 convolution kernels to combine features between different layers with dense connectivity.The student model is a simple end-to-end architecture with five convolutional layers and a fully-connected layer.The student model is trained to fit the teacher model by knowledge distillation.Furthermore,according to the values of BN scaling factors towards zero to identify insignificant channels,the paper prunes redundant channels for student network,yielding a compact model with comparable accuracy.The experimental results on GTSRB and BTSC datasets show the recognition accuracy and less parameters of the student model have great advantages compared with teacher model and other popular models.