基于YOLOv5s的改进实时红外小目标检测  

Improved real-time infrared small target detection based on YOLOv5s

作　　者：谷雨[1] 张宏宇 彭冬亮[1] GU Yu;ZHANG Hong-yu;PENG Dong-liang(School of Automation,Hangzhou Dianzi University,Hangzhou 310018,China;HDU-ITMO Joint Institute,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,浙江杭州310018 [2]杭州电子科技大学圣光机联合学院,浙江杭州310018

出　　处：《激光与红外》2024年第2期281-288,共8页Laser & Infrared

基　　金：浙江省自然科学基金项目(No.LZ23F030002)资助。

摘　　要：针对红外图像分辨率低、背景复杂、目标细节特征缺失等问题,提出了一种基于YOLOv5s的改进实时红外小目标检测模型Infrared-YOLOv5s。在特征提取阶段,采用SPD-Conv进行下采样,将特征图切分为特征子图并按通道拼接,避免了多尺度特征提取过程中下采样导致的特征丢失情况,设计了一种基于空洞卷积的改进空间金字塔池化模块,通过对具有不同感受野的特征进行融合来提高特征提取能力;在特征融合阶段,引入由深到浅的注意力模块,将深层特征语义特征嵌入到浅层空间特征中,增强浅层特征的表达能力;在预测阶段,裁减了网络中针对大目标检测的特征提取层、融合层及预测层,降低模型大小的同时提高了实时性。首先通过消融实验验证了提出各模块的有效性,实验结果表明,改进模型在SIRST数据集上平均精度均值达到了95.4%,较原始YOLOv5s提高了2.3%,且模型大小降低了72.9%,仅为4.5 M,在Nvidia Xavier上推理速度达到28 f/s,利于实际的部署和应用。在Infrared-PV数据集上的迁移实验进一步验证了改进算法的有效性。提出的改进模型在提高红外图像小目标检测性能的同时,能够满足实时性要求,因而适用于红外图像小目标实时检测任务。In this paper,an improved infrared small target detection model,infrared-YOLOv5s,based on YOLOv5s is proposed to address the problems of low resolution,complex background and lack of detailed features of infrared images.In feature extraction stage,SPD-Conv is used for down-sampling,which divides the feature map into feature sub-maps and concatenate them by channel to avoid the loss of features caused by down-sampling in the process of multi-scale feature extraction.And an improved atrous spatial pyramid pooling module is designed to improve feature extraction capabilities by fusing features with different receptive fields.Then,in feature fusion stage,a deep-to-shallow attention module is introduced to embed deep semantic features into shallow spatial features to enhance the expression of shallow features.Moreover,in prediction stage,the prediction layers,feature extraction layers and feature fusion layers for large target detection in the network are cut down to reduce the model size and improve real-time performance at the same time.The effectiveness of each module is verified by ablation experiments,and experimental results show that the proposed model achieves 95.4%mAP 0.5 of on SIRST dataset,which is 2.3%higher than that of original YOLOv5s.The model size is reduced by 72.9%to 4.5 MB,and the inference speed on Nvidia Xavier reaches 28 f/s,which is conducive to the actual deployment and application.Therefore,the effectiveness of the proposed model is further verified by transfer experiments using Infrared-PV dataset,and the proposed model can meet the real-time requirements while improving the performance of small target detection in infrared images,and is suitable for the task of real-time small target detection in infrared images.

关 键 词：红外小目标检测 YOLOv5s 注意力机制 特征融合 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]