一种基于SAM-MSFF网络的低照度目标检测方法  

A Low-Light Object Detection Method Based on SAM-MSFF Network

作　　者：江泽涛[1] 李慧[1] 雷晓春[1] 朱玲红[2] 施道权 翟丰硕 JIANG Ze-tao;LI Hui;LEI Xiao-chun;ZHU Ling-hong;Shi Dao-quan;ZHAI Feng-shuo(The Key Laboratory of Image and Graphic Intelligent Processing in Guangxi,Guilin University of Electronic Technology,Guilin,Guangxi 541004,China;Nanchang Hangkong University,Nanchang,Jiangxi 330063,China)

机构地区：[1]桂林电子科技大学广西图像图形与智能处理重点实验室,广西桂林541004 [2]南昌航空大学,江西南昌330063

出　　处：《电子学报》2024年第1期81-93,共13页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62172118,No.61876049);广西自然学科基金(No.2021GXNSFDA196002);广西图像图形智能处理重点实验项目(No.GIIP2006,No.GIIP2007,No.GIIP2008);广西研究生教育创新计划项目(No.YCB2021070,No.YCBZ2018052,No.YCSW2022269,No.2021YCXS071)。

摘　　要：由于低照度图像具有对比度低、细节丢失严重、噪声大等缺点,现有的目标检测算法对低照度图像的检测效果不理想.为此,本文提出一种结合空间感知注意力机制和多尺度特征融合(Spatial-aware Attention Mechanism and Multi-Scale Feature Fusion,SAM-MSFF)的低照度目标检测方法 .该方法首先通过多尺度交互内存金字塔融合多尺度特征,增强低照度图像特征中的有效信息,并设置内存向量存储样本的特征,捕获样本之间的潜在关联性;然后,引入空间感知注意力机制获取特征在空间域的长距离上下文信息和局部信息,从而增强低照度图像中的目标特征,抑制背景信息和噪声的干扰;最后,利用多感受野增强模块扩张特征的感受野,对具有不同感受野的特征进行分组重加权计算,使检测网络根据输入的多尺度信息自适应地调整感受野的大小.在ExDark数据集上进行实验,本文方法的平均精度(mean Average Precision,mAP)达到77.04%,比现有的主流目标检测方法提高2.6%~14.34%.The existing object detection methods are insufficient for low-light images due to their intrinsic property such as low contrast,detail loss and high noise.To solve this problem,a low-light object detection method that combines spatial-aware attention mechanism with multi-scale feature fusion(SAM-MSFF)is proposed.Firstly,multi-scale features are fused by multi-scale interactive memory pyramid to enhance effective information under low-illumination condition,and features of memory vector storage samples are set to capture potential correlation between samples.Then,a spatial-aware attention mechanism is introduced to obtain long-distance context information and local information of features in spatial domain,thereby enhancing the object features in low-light images and suppressing the interference of background in-formation and noise.Finally,multiple receptive field enhancement module is used to expand receptive field of the features,and the features with different receptive fields are grouped and re-weighted,so that detection network can adaptively adjust the size of receptive field according to input multi-scale information.Experimental results on the ExDark dataset show that mAP(mean Average Precision)of the proposed method reaches 77.04%,which is 2.6%~14.34%higher than existing main-stream object detection methods.

关 键 词：低照度图像 目标检测 空间感知注意力机制 多尺度特征融合 多感受野增强模块 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]