深度可分离卷积网络的驾驶状态识别算法

针对嵌入式设备内存小及多分类准确率低等导致驾驶员检测问题,提出经过深度可分离卷积网络改进而成的,快速下采样网络(fast downsampling network,MF-Net)驾驶状态识别系统.即将快速下采样策略应用于深度可分离卷积网络,在12层内执行32倍下采样,以有效降低计算成本、增加信息容量并实现性能改进.实验结果表明:与VGG(visual geometry group)和ResNet 50等其他卷积神经网络(CNN)模型相比,所提出的MF-Net模型深度可分离卷积大大减少参数量,快速下采样方案的运用增加了网络的信息容量,不仅模型较小且在驾驶员状态分类方面能够表现出更好的性能.同时,信息容量的增加可以对更多信息进行编码,加深对图像内容的理解,有利于之后的嵌入式系统移植.

Driving State Recognition Algorithm Based on Deep Separable Convolutional Network

Aiming at the problem of driver detection caused by the small memory of embedded devices and the low accuracy of multi-classification, a fast fast down sampling network(MF-Net)driving state recognition system improved by deep separable convolutional network is proposed, the key idea is applying a fast downsampling strategy to deep separable convolutional networks, which performs 32-fold downsampling within 12 layers to reduce computational cost seffectively, increase information capacity, and achieve performance improvements. The experimental results show that: compared with other convolutional neural network(CNN)models such as VGG(visual geometry group)and ResNet 50, MF-Net model has deep separable convolutions that reduce the amount of parameters greatly and the application of fast down sampling increase the information capacity of the network, the model not only smaller but also show better performance in the classification of driver status, at the same time, the increase in information capacity can encode more information and deepen the understanding of the image content, which is beneficial to transplantation of embedded systemsin the future.