基于混合域注意力机制的鞋印检索算法

Shoeprint Retrieval Algorithm Based on Hybrid Domain Attention Mechanism

鞋印图像识别是计算机视觉在公安一线工作中的一项重要应用。当前公安侦查工作中鞋印图像无法进行精准识别的问题制约了工作效率与质量的提高,归纳起来主要是囿于鞋印现场提取的复杂情况、鞋印花纹图样的复杂特征以及鞋印图像的残缺不清。针对残缺鞋印,提出一种在经典ResNet34网络中嵌入基于混合域双重注意力机制的鞋印检索算法。算法通过通道注意力模块在各个通道中产生不同的信号特征,借助空间注意力模块将原始特征信息中的关键信息映射至另一空间;融合ResNet34网络中layer2和layer4的卷积层特征,低层卷积层保留图片的细节特征,高层卷积层保留语义特征,将聚合卷积层特征作为特征描述符,使图片特征信息更加丰富。实验结果表明,该算法在CSS-200数据集中top1的准确率达到了56%;在part-FID数据集上top1%达到了45.32%,准确度提高了2.87%。

Shoeprint image recognition is an important application of computer vision in the frontline work of public security.In current public security investigations,the inability to accurately identify shoeprint images restricts the improvement of work efficiency and quality,mainly due to the complex situation of shoe print site extraction,the complex features of shoe print patterns and the incomplete shoe print images.Aiming at the incomplete shoe print,a shoeprint retrieval algorithm based on the hybrid domain dual attention mechanism embedded in the classic ResNet 34 network was proposed.The algorithm generated different signal features in each channel through the channel attention module,and mapped the key information in the original feature information to another space with the help of the spatial attention module.The algorithm merged the convolutional layer features of layer2 and layer4 in the ResNet 34 network.The low-level convolutional layer retained the detailed features of the image,the high-level convolutional layer retained the semantic features,and the aggregated convolutional layer features were used as feature descriptors to enrich the image feature information.The experimental results showed that the accuracy of the algorithm on the top1 in the CSS-200 data set and on the top1%in the part-FID data set reached 56%and 45.32%,respectively.And the accuracy was improved by 2.87%.