改进深度可分离卷积的SSD车型识别

针对现有车辆识别方法对于车型实时识别能力不足的问题,提出一种改进的深度可分离卷积的SSD(single shot multibox detector)算法用于车型识别研究.首先,利用深度可分离卷积网络进行特征提取,并引入反残差模块来解决因通道数少、特征压缩导致的准确率下降问题.其次,以车辆的刚体特性为依据,重新设计区域候选框,减少模型参数运算量.最后,在BIT-Vehicle数据集上进行消融实验来对比不同网络模型性能差异.结果表明:改进的深度可分离卷积的SSD车型识别方法有更好的车型识别效果,可以达到96.12％的识别精度,检测速度提高至0.078 s/帧.

Vehicle type recognition based on improved depthwise separable convolution SSD

Aiming at the problem of insufficient real-time recognition capabilities of existing vehicle recognition methods, a single shot multibox detector(SSD) algorithm based on improved depthwise separable convolution is proposed for vehicle type recognition. Firstly, this paper proposes to extract the features using depthwise separable convolution network, and introduces the inverted residuals module to solve the problem of reduced accuracy due to the small number of channels and feature compression. Secondly, based on the rigid body characteristics of the vehicles, the region candidate frame is redesigned to reduce the amount of model parameter calculation. Finally, ablation experiments are performed on the BIT-Vehicle dataset to compare the performance differences of different network models. The results show that the improved depthwise separable convolution SSD vehicle type recognition method can achieve a recognition accuracy of 96.12%, and the detection speed increases to 0.078 s/frame.