基于卷积神经网络的交通标志识别

Traffic Signs Recognition Based on Convolution Neural Network

实际的交通场景中,交通标志图像通常会受到运动模糊、背景干扰、形状畸变等因素的影响,快速准确的识别交通标志具有相当大的难度。传统卷积神经网络在图像识别过程中,池化等操作会导致图像细节信息丢失,影响交通标志识别的准确率。针对这一问题,提出一种低压缩度特征卷积神经网络模型,通过在全连接层聚合压缩程度较低的特征图,实现对图片细节特征的表达。实验结果表明,和传统的卷积神经网络相比,该模型具有更高的识别率。

In actual traffic scenes, traffic sign images are usually affected by factors such as motion blur, background interference, shape distortion, etc. Therefore, it is quite difficult to quickly and accurately identify traffic signs. Pooling can cause the loss of image details in the process of image recognition based on traditional convolution neural network, and it affects the accuracy of traffic sign recognition. Aiming at this problem, convolution neural network based on low compression degree feature is proposed to express the detailed feature by aggregating low compression degree feature in fully connected layers. The experiment results show that compared with the traditional convolutional neural network, the proposed model has a higher recognition rate.