摘要
针对带钢材料表面缺陷检测中感受野受限导致目标漏检率高的问题,基于YoloX-s模型提出一种膨胀卷积与注意力机制融合的目标检测算法。在Backbone部分采用SPPF结构替换SPP结构,在Neck部分引入混合膨胀卷积模块用以增大检测的感受野,嵌入注意力机制ECA-net模块,保留特征图更多的通道信息,减少漏检率。后处理阶段采用CIoU损失函数,提高模型召回率。实验结果表明,改进算法在NEU-DET数据集上的mAP达到80.8%,较原模型提高4.6%,检测速度达到160 f/s,在带钢材料表面缺陷检测中具有一定的使用价值。
Aiming at the problem of high target missed detection rate caused by limited receptive field in surface defect detection of strip materials,an object detection algorithm based on model named Exceeding YOLO Series-s(YoloX-s)was proposed by combining expansion convolution and attention mechanism.In the Backbone part,the spatial pyramid pooling-fast(SPPF)structure was used to replace the spatial pyramid pooling(SPP)structure.In the Neck part,the hybrid expansion convolution module was introduced to increase the receptor field of detection,and the attention mechanism named efficient channel attention for deep convolutional neural networks(ECA-net)module was embedded to retain more channel information of the feature map and reduce the missed detection rate.The Complete-IoU(CIoU)loss function was adopted in the post-processing stage,which made the detection box closer to the real box and improved the model recall.Experimental results show that the mAP of the improved algorithm on the NEU surface defect(NEU-DET)dataset reaches 80.8%,which is 4.6%higher than that of the original model,and the detection speed reaches 160 f/s,which has certain use value in the surface defect detection of strip materials.
作者
曹义亲
曹鑫晨
CAO Yi-qin;CAO Xin-chen(School of Software,East China Jiaotong University,Nanchang 330013,China)
出处
《计算机工程与设计》
北大核心
2024年第11期3312-3319,共8页
Computer Engineering and Design
基金
国家自然科学基金项目(61861016)
江西省科技支撑计划重点基金项目(20161BBE50081)。
