CoT-YOLO水下目标检测算法  

CoT-YOLO underwater target detection algorithm

作　　者：苏佳[1] 冯康康 梁奔 侯卫民[1] SU Jia;FENG Kang-kang;LIANG Ben;HOU Wei-min(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China)

机构地区：[1]河北科技大学信息科学与工程学院,河北石家庄050018

出　　处：《计算机工程与设计》2024年第7期2119-2126,共8页Computer Engineering and Design

基　　金：国家自然科学基金青年科学基金项目(62105093);装备预研重点实验室基金项目(6142A010301);2023年河北科技大学创新基金项目(XJCXZZSS202303)。

摘　　要：水下检测由于背景复杂、光线暗淡、目标遮挡重叠等问题导致检测精度较低,提出一种CoT-YOLO水下目标检测算法提高检测精度。使用YOLOv5s作为基础模型,构建注重上下文信息的卷积神经网络,充分利用特征信息,增强全局特征表达能力,解决模型漏检、误检问题;改用解耦头,加快收敛速度;增添新的检测层并重获先验框,增强模型对小目标的检测能力,提高水下小目标检测效果;采用EIoU损失函数提高目标边界框的定位与回归。实验结果表明,改进后算法精确度达到77.9%,相较于基线提升了3.7%,mAP提升了5.2%,验证了该方法的有效性。Underwater detection has low detection accuracy due to complex background,dim light and overlapping target occlusion.The CoT-YOLO underwater target detection algorithm was proposed to improve detection accuracy.Using YOLOv5s as the base model,a convolutional neural network focusing on contextual information was constructed to make full use of feature information and enhance global feature representation to solve the problem of missed and false detection of the model.The decoupling head was used instead to speed up convergence.New detection layers were added and priori frames were regained to enhance the detection capability of the model for small targets and improve the detection effect of small underwater targets.The EIoU loss function was used to improve the target bounding box localization and regression.Experimental results show that the improved algorithm achieves 77.9%accuracy,a 3.7%improvement compared to the baseline,and a 5.2%improvement in mAP,verifying the effectiveness of the method.

关 键 词：目标检测 YOLOv5 卷积神经网络 特征信息 分类回归 解耦头 EIoU损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]