基于深度哈希网络的车型识别方法

A Vehicle Recognition Algorithm Based on Deep Hashing Network

针对车型识别任务的特点,设计了一种基于深度哈希网络的车型识别方法,实现了在类间差异不明显、样本量较少的情况下进行车型检索和分类。对数据增广方法进行研究,针对车型数据集的特点,提出了适用于车型识别的数据增广方法,有效提升了小样本车型识别的准确率。深度哈希网络采用改进的HashNet网络来快速学习车辆的二值特征表达,针对深度哈希网络使用全连接层导致参数过多的问题,提出了HashNet-GAP网络,以全局平均池化层替换了HashNet中的部分全连接层。相对于HashNet网络,大幅度减少了参数数量,提升了前向计算速度和网络性能。实验结果表明,该车型识别方法能够对类间差距很小的不同车型进行有效识别,在小样本数据集上取得80.0%的Top1准确率,并且能够显著降低模型的存储消耗和内存消耗。

According to the characteristics of vehicle recognition tasks,a vehicle recognition method based on deep hashing network is designed,which realizes the retrieval and classification of vehicles when the differences between the classes are not obvious and the size of training set is small.The data augmentation method is studied.According to the characteristics of the vehicle data set,a data augmentation method suitable for vehicle recognition is proposed,which effectively improves the accuracy of small sample recognition.The deep hashing network uses the improved HashNet to quickly learn the binary feature descriptors of the vehicle.The HashNet-GAP network is proposed to reduce the number of parameters caused by full connected layers.The full connected layer is partially replaced by the global average pooling layer.Compared with the HashNet,the number of parameters is greatly reduced,and the forward calculation speed and network performance are improved.The experiment shows that the vehicle recognition method can effectively recognize different vehicle models with small gaps between classes,achieve 80.0%Top1 accuracy on a small dataset,and can significantly reduce storage consumption and memory consumption of the network model.