基于三维注意力机制的车辆重识别算法

Vehicle Recognition Algorithm Based on 3D Attention Mechanism

摘要:

为解决套牌车识别难度大的问题,利用深度学习的技术,基于ResNet-50,结合通道注意力机制和位置注意力机制,设计了一种三维注意力机制对近似车辆进行精确识别;当前大部分注意力算法都关注于一维的通道注意力和二维的位置注意力,而处理的图像是三维的,这些注意力机制不能将注意力集中在所有需要关注的区域,造成部分关键信息遗失;该三维注意力机制在多种视觉任务下均有很好的效果,在Cifar100数据集上,相比SENet有1.12%的提升,在PKU VehicleID数据集上,相比SENet平均有2%的提升。

Abstract:

In order to solve the problem of the difficulty in recognizing the fake-licensed vehicles, this paper uses deep learning technology and combines channel attention mechanism and position attention mechanism based on RESNET-50 to design a three-dimensional attention mechanism for accurate identification of similar vehicles. At present, most attention algorithms focus on one-dimensional channel attention and two-dimensional positional attention, while images are processed in three dimensions. Therefore, these attention mechanisms cannot focus on all areas requiring attention, resulting in the loss of some key information. The three-dimensional attentional mechanism works well in a variety of visual tasks, with a 1.12 per cent improvement over SENet on the Cifar100 dataset, and a 2 per cent improvement over SENet on average on the PKU VehicleID dataset.