【摘要】 针对同时存在多种小目标、大长宽比目标缺陷的锂电池极片复杂表面，基于可变形卷积和YOLOv5提出DDCNet-YOLO算法模型.在主干网络部分构建出可变形下采样卷积主干网络(DDCNet)，在特征融合部分引入上下文增强模块(CAM)，并使用构造的可变形卷积块(DCB)替换C3模块，在检测头部分设计带有注意力机制的解耦头AD-Head.提出RIoU方法优化不同长宽比目标的损失计算.实验表明，DDCNet-YOLO模型相较于YOLOv5s及YOLOv5m模型在mAP50上分别提高了6.2个百分点和3.7个百分点.仅通过DDCNet和注意力机制解耦头构建了DDCNet-YOLOs轻量化模型，与YOLOv5s模型相比，参数量减少7.2个百分点，mAP50∶95提升8.9个百分点.对2种模型通过C++的方式进行了部署.本研究所提出的2种算法模型分别侧重于精度和轻量化，都能够在满足一定实际检测速度的条件下，达到较高的检测精度.更多还原

【Abstract】 The DDCNet-YOLO algorithm model was proposed based on the deformable convolution and YOLOv5,aiming at the complex lithium battery electrode surface with multiple small object defects and large aspect ratio object defects at the same time. The deformable downsampling convolution network(DDCNet) was constructed in the backbone. The context augmentation module(CAM) was introduced in the feature fusion part and the deformable convolution block(DCB) was used to replace the C3 module. AD-Head, a decoupling head with an attention mechanism, was designed in the head part. The RIoU method was proposed to optimize the loss calculation for different aspect ratio objects. Experiments showed that the DDCNet-YOLO model improved the mAP50 by 6.2percentage points compared to YOLOv5s model and by 3.7 percentage points compared to YOLOv5m model. The lightweight model DDCNet-YOLOs, constructed by DDCNet and a decoupling head with an attention mechanism.The DDCNet-YOLOs improved the mAP50:95 by 8.9 percentage points and reduced the number of parameters by7.2 percentage points, compared with the YOLOv5s model. In addition, both models were deployed based on the C++. The two algorithmic models focus on accuracy and speed respectively, but both can achieve high accuracy under the condition of meeting the actual detection speed requirement.更多还原