用于交通图像识别的改进尺度依赖池化模型

摘要

针对交通标志在自然场景中所占的比例较小、提取的特征量不足、识别准确率低的问题,提出改进的尺度依赖池化(SDP)模型用于小尺度交通图像的识别.首先,基于神经网络深卷积层具有较好的轮廓信息与类别特征,在SDP模型只提取浅卷积层特征信息的基础上,使用深卷积层特征补足型SDP(SD-SDP)映射输出,丰富特征信息;其次,因SDP算法中的单层空间金字塔池化损失边缘信息,使用多尺度滑窗池化(MSP)将特征池化到固定维度,增强小目标的边缘信息;最后,将改进的尺度依赖池化模型应用于交通标志的识别.实验结果表明,与原SDP算法比较,提取特征量增加,小尺度交通图像的识别准确率较好地提升.%Aiming at these problems that the traffic sign has a small proportion in the natural scene,the extracted features are insufficient and the recognition accuracy is low,an improved Scale Dependent Pooling (SDP) model was proposed for the recognition of small-scale traffic images.Firstly,because the deep convolution layer of neural network has better contour information and class characteristics,Supplementary Deep convolution layer characteristic Scale-Dependent Pooling (SD-SDP) model for deep convolution layer characteristic was used to extract features based on the feature information of shallow convolution by SDP model,enriching feature information.Secondly,the Multi-scale Sliding window Pooling (MSP) was used to make up the edge information of the target object,instead of the single-layer spatial pyramid method in the original SDP algorithm.Finally,the improved SDP model was applied to the recognition of traffic signs.The experimental result show that,compared to SDP algorithms,the extracted feature dimension increases and the accuracy of small scale traffic image recognition is improved.