基于注意力机制与多尺度特征融合的电极缺陷YOLO检测算法

为了满足锂离子电池电极缺陷检测精度与实时性的需求,解决电极图像背景噪声复杂、缺陷微小且对比度低等问题,提出一种基于注意力机制与多尺度特征融合的电极缺陷YOLO检测算法.在YOLOv4的基础上,首先,将SE(squeeze-and-excitation)注意力模块嵌入特征提取主干网络中,区分feature map中不同通道的重要性,强化目标区域的关键特征,提高网络的检测精度;其次,加入融合空洞卷积的池化金字塔(ASPP)结构,增大网络感受野的同时最大程度地保留多尺度特征信息,提高算法对小目标的检测性能;然后,设计一种多尺度稠密特征金字塔,在三尺度特征图的基础上增加一个浅层特征,采用稠密连接的方式融合特征,提升浅层细节特征与高级语义信息的融合能力,增强对微小缺陷特征的提取;最后,采用$ K $-means++算法聚类先验框,引入focal loss损失函数增大小目标样本的损失权重,有效提高网络学习的收敛速度.实验结果表明,所提算法较原YOLOv4模型的mAP值提升6.42%,较其他常用算法综合性能上有着较大的优势,可较好地满足实际工业生产的实时监测需求.

Electrode defect YOLO detection algorithm based on attention mechanism and multi-scale feature fusion

In order to meet the requirements of detection accuracy and real-time performance of lithium-ion battery electrode defects, and to solve the problems of complex background noise of electrode images, small defects and low contrast, this paper proposes a electrode defect YOLO detection algorithm based on attention mechanism and multi-scale feature fusion. On the basis of YOLOv4. First, we embed the SE(squeeze-and-excitation) attention module in the feature extraction backbone network to distinguish the importance of different channels in the feature map, strengthen the key features of the target area, and improve the detection accuracy of the network; secondly add the pooling pyramid(ASPP) structure fused with atrous convolution to increase the network receptive field while retaining the multi-scale feature information to the greatest extent, and improve the detection performance of the algorithm for small targets; then design a multi-scale dense feature pyramid, on the basis of the three-scale feature map, a shallow feature is added, and the feature is fused by dense connection, which improves the fusion ability of shallow detail features and high-level semantic information, and enhances the extraction of small defect features; finally, the K-means++ algorithm is used for clustering. In the prior box, the focal loss function is introduced to increase the loss weight of small target samples, which effectively improves the convergence speed of network learning. The experimental results show that the mAP value of the proposed algorithm is increased by 6.42% compared with the original YOLOv4 model, which has a greater advantage in comprehensive performance than other commonly used algorithms, and can better meet the real-time monitoring needs of actual industrial production.