一种改进YOLOv5的小目标检测算法

An improved algorithm for small object detection based on YOLOv5

随着自动驾驶等人工智能技术的日渐成熟,对更快、更准确的小目标检测算法的需求也在增加。为此,提出一种基于单阶段检测算法YOLOv5的小目标检测改进算法,以提高在检测较小对象方面的性能。为了实现目标,将研究替换原始模型中的部分结构和因素会对检测性能和推理时间产生的影响。为此,提出了一系列不同尺度的模型并将其命名为“YOLOv5-Sobj”,当以50%的IoU检测较小的物体时,这些模型的mAP提升高达5.3%,而代价是与原始YOLOv5对比,推理时间增加了3 ms。实验目的在于研究提高小目标检测效果的相应方法。

As artificial intelligence technologies such as autonomous driving mature,so does the demand for faster and more accurate small object detection algorithms.To this end,an improved small object detection algorithm is proposed based on the single-stage detection algorithm YOLOv5 to improve the performance in detecting smaller objects.In order to achieve the goal,research will be conducted on the impact of replacing partial structures and factors in the original model on detection performance and inference time.To this end,a series of models with different scales were proposed and named“YOLOv5-Sobj”.When detecting smaller objects with 50%IoU,compared with the original YOLOv5,the mAP improvement of these models is as high as 5.3%,while the inference time increases by 3 ms.The research in this paper aims to provide insights into how certain changes affect small object detection.