在内窥镜手术过程中，外科医师需实时掌握手术器械的位置信息。现有目标检测算法受反光和阴影等因素影响，其准确度和漏检率仍有优化的空间。本文提出一种基于改进YOLOv5s的手术器械检测与分割方法。首先，通过Gamma校正算法校正图像的亮度和对比度，以解决手术器械的反光和阴影遮挡等问题；其次，设计CBAM和动态卷积模块，增加重要特征信息的权重，以进一步提高目标检测的准确度并减少模型的漏检率；同时，优化空间金字塔池化模块以扩大感受野，从而更好地识别多尺度目标；最后，设计FPN语义分割头，以实现语义分割功能。在内窥镜手术数据集上的实验结果表明，本文目标检测的mAP@0.5为98.2%，语义分割的mIoU为94.0%。所提方法可辅助外科医师快速掌握手术器械的位置和类型，提高手术效率。

In the process of endoscopic surgery, surgeons need to know the position information of surgical instruments in real time. The existing target detection algorithms are affected by factors such as reflection and shadow, and there is still optimization space for the accuracy and missed detection rate. This paper proposes a detection and segmentation method of surgical instruments based on improved YOLOv5s. Firstly, the brightness and contrast of images are corrected by Gamma correction algorithm to solve the problems of reflection and shadow occlusion of surgical instruments. Secondly, convolutional block attention module(CBAM) and dynamic convolution module are designed to increase the weight of important feature information, which further improves the accuracy of target detection and reduces the missed detection rate of the model. At the same time, the spatial pyramid pooling module is optimized to expand the receptive field, so as to better identify multi-scale targets. Finally, the feature pyramid networks (FPN) semantic segmentation head is designed to realize the semantic segmentation. Experimental results on endoscopic surgery dataset show that the mAP@0.5 of target detection in this paper is 98.2%, and the mIoU of semantic segmentation is 94.0%. The proposed method can assist surgeons to quickly grasp the position and type of surgical instruments, and improve the efficiency of surgery.