针对目前胶囊内窥镜病灶检测模型存在检测疾病单一且效率低等问题，提出了一种基于YOLOv5的胶囊内窥镜病灶区域检测方法。该方法在原始YOLOv5基础上进行了如下改进：首先，在主干网络Backbone部分，添加一个CBAM（convolutional block attention module）模块，增强模型对重要特征的突出能力；其次，在头部网络Head部分，添加一个检测头，增强模型对小目标的检测能力；最后，将原始YOLOv5的泛化交并比（generalized intersection over union, GIoU）损失函数替换成完整交并比（complete intersection over union, CIoU）损失函数，使模型训练时更快地收敛。本文提出的方法在长江大学第一临床医学院提供的胶囊内窥镜影像数据上进行了实验，精确率达到了93.6%，召回率达到了94.3%，mAP@0.5达到了97.2%，而且检测速度达到了每帧0.0272s。实验结果表明提出的方法是有效的、灵活的、鲁棒的，能够满足临床医学诊断的实际需求。

Since the existing capsule endoscopy detection model has the problem of single lesion detection and low efficiency, a novel lesion region detection method of capsule endoscopic based on the improved YOLOv5 was proposed. The original YOLOv5 had been improved in the following three aspects. Firstly, the convolutional block attention module (CBAM) was incorporated into the Backbone of YOLOv5, which was used to enhance the presentation ability of the target feature. Secondly, a detection head was added to the Head of YOLOv5, which was used to enhance the detection ability of the small targets. Finally, the GIOU loss function of YOLOv5 was replaced with the CIOU loss function, which leaded to quicker convergence during training. The proposed method was tested on the capsule endoscopic image data provided by the First Clinical Medical College of Yangtze University. It's accuracy rate reached 93.6%, the recall rate reached 94.3%, the detection rate reached 97.2% on mAP@0.5 and a detection speed of 0.027 2 s per frame for lesion area detection on the dataset. The experimental results show that the proposed method is effective, flexible, and robust, which can meet the practical needs of clinical medical diagnosis.