Abstract: Object detection is a research hotspot in the current industry and academia, and with the further development of deep learning and related needs, more accurate detection models are needed. YOLO, a new object detection method that emphasizes single-stage models. By studying the traditional object detection network model, this paper proposes a plug-and-play self-calibrating convolutional module. Based on the existing model YOLOv5, this paper inserts this module into the C3 module of the YOLOv5 backbone network. After comparative experiments, it is found that the curve graph, detection map and heat map generated by the two models show that the convolution module can effectively improve the target detection effect of the original YOLOv5 model.

Key words: Object detection; YOLOv5; Convolutional neural networks; Self-calibrating convolution

摘要： 目标检测是当前工业界、学术界的研究热点，随着深度学习以及相关需求的进一步发 展，需要更精准的检测模型。YOLO，一种新的对象检测方法，强调的是单阶段的模型。文 章通过研究传统目标检测网络模型，提出一种即插即用的自校准卷积模块。文章基于现有模 型 YOLOv5，将该模块插入 YOLOv5 主干网络的 C3 模块中。经过对比实验后发现，两种模 型所生成的曲线图、检测图、热力图均表明，该卷积模块能够有效提升 YOLOv5 原模型的目 标检测效果。

关键词: 目标检测；YOLOv5；卷积神经网络；自校准卷积