Abstract: Considering the problems of how to improve the performance of IR object detection in complex scenes and how to balance the lightness and accuracy of the algorithm, this paper proposes a lightweight IR target detection algorithm based on YOLOv5. This paper designs a Ghost feature extraction module that incorporates parallel convolution to reduce the complexity of the model and speed up the inference by using the idea of structural re-referencing to simplify the model at the time of inference. The network is pruned to add small target detection layer and improve the detection performance in complex scenes. Secondly, a Shuffle Ghost module is designed in the feature fusion module to shuffle ghost features and standard convolutional features to minimize the negative impact of ghost features on network performance. Finally, a decoupled detection head module is designed to decompose the classification and detection tasks to improve the localization and detection capability of the network in complex environments. The experimental results show that proposed algorithm has a 4.7 percentage points improvement in mAP, a 42.7% reduction in the parameters, and an 8.5% reduction in inference latency compared to YOLOv5s, and is able to achieve an ideal balance of detection performance and lightweight deployment.

Key words: IR object detection, YOLOv5, ghost net, structural re-referencing

摘要： 针对如何提高在复杂场景下红外目标检测性能以及如何平衡算法的轻量化与精确度等问题，提出了一种基于YOLOv5的轻量化红外目标检测算法。该算法设计一种融合并行卷积的Ghost特征提取模块，利用结构重参化的思想，在推理时将模型简化，降低模型的复杂度且加快推理速度；同时对网络进行剪枝，增加了小目标检测层，提高了模型对复杂场景的检测性能；在特征融合模块设计了一种混洗Ghost模块，将Ghost特征和标准卷积特征进行混洗，尽可能减少Ghost特征对网络性能的负面影响；设计了一个解耦检测头模块，将分类与检测任务进行分解，提高了网络在复杂环境下的定位检测能力。实验结果表明，与YOLOv5s相比，提出的算法mAP提高了4.7个百分点，参数量降低了42.7%，推理延时减少了8.5%，能够在检测性能和轻量化上达到理想平衡。

关键词: 红外目标检测, YOLOv5, 幽灵网络, 结构重参化