【摘要】 针对当前复杂卡口场景下对车型细粒度识别存在数据集小和特征差异小导致识别精度不高的问题,提出一种基于分类学习和度量学习的多任务学习的卷积神经网络,用于实现细粒度车型的识别。具体而言,首先通过迁移学习利用具有一定相似性的数据中训练好的优良模型,将其中可用的知识迁移出来,在小样本数据集中微调。然后基于多任务学习的思想,结合分类学习和度量学习设计双通道模型,进一步约束参数的学习。实验结果表明,通过迁移学习并且设置损失函数约束来提供细粒度信息的双通道模型,对于小数据卡口场景下的车型识别效果显著提高。更多还原

【Abstract】 Aiming at the problem that the fine-grained identification of the model in the current complex bayonet scenario has small data set and small difference in features, resulting in low recognition accuracy. Proposes a multi-task learning convolutional neural network based on classification learning and metric learning, which is used to realize the identification of fine-grained models. Specifically, the method first learns to use the well-trained model in the data with certain similarity through transfer learning, to migrate the knowledge available and fine-tune it in the small sample data set. Then based on the idea of multi-task learning, combined with classification learning and metric learning to design a two-channel model to further constrain the learning of parameters. The experimental results show that the two-channel model of fine-grained information is provided by transfer learning and setting loss function constraints, and the recognition effect of the vehicle model in the small data bayonet scenario is significantly improved.更多还原