Abstract: A target detection network based on the improvement of YOLOv5s is proposed to solve the problem of low detection efficiency due to the shortcomings of vehicle-mounted infrared image, such as inconspicuous performance of detail information, low contrast, and poor imaging effect. Firstly, a dynamic detection head based on an attention mechanism is added to the head network, with an attention mechanism between feature layers for scale perception, an attention mechanism between spatial positions for spatial perception, and an attention mechanism within the output channel for task perception, which makes the network focus more on detecting the foreground target associated with the task, and improves the expressive ability of the model target detection head. Then the CIOU bounding box loss function is replaced by MPDIOU during training to improve the localization accuracy and efficiency of the model. Finally, the lightweight network FasterNet is added to the C3 module in the end of the neck network to improve the real-time performance of the model. The experimental results show that the improved network model improves the mAP by 2.1% compared with the original network model before the improvement, and the weight size of the model is almost unchanged. It meets the requirements of small size and real-time, and is suitable for in-vehicle embedded systems.

Key words: Target detection; YOLOv5s; Attention mechanism; Loss function; Vehicle-mounted infrared image

摘要： 针对车载红外图像细节信息表现不明显、对比度低、成像效果差等缺点导致检测效率 不高的问题，文章提出了一种基于 YOLOv5s 改进的目标检测网络。首先在头部网络中添加一 个基于注意力机制的动态探测头，其特征层间的注意力机制用于尺度感知，空间位置间的注 意力机制用于空间感知，输出通道内的注意力机制用于任务感知，这使网络更加重点关注检 测任务中相关联的前景目标，提升模型目标检测头的表达能力。然后在训练时用 MPDIOU 替 换 CIOU 边界框损失函数，提升模型的定位精度与效率。最后把轻量级网络 FasterNet 添加到 颈部网络末端中的 C3 模块，提升模型的实时性。实验结果表明，改进后的网络模型较改进前 原始网络模型的 mAP 提升了 2.1%，模型权重大小几乎不变，满足体积小与实时性的需求， 适用于车载嵌入式系统中。

关键词: 目标检测；YOLOv5s；注意力机制；损失函数；车载红外图像