并行多特征提取网络的红外图像增强方法  

Parallel multifeature extracting network for infrared image enhancement

作　　者：庞忠祥 刘勰 刘桂华[1] 龚泿军 周晗 罗洪伟 Pang Zhongxiang;Liu Xie;Liu Guihua;Gong Yinjun;Zhou Han;Luo Hongwei(School of Information Engineering,Southwest University of Science and Technology,Mianyang 621000,China;Shenzhen Launch Digital Technology Co.,Ltd,Shenzhen 518000,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621000 [2]深圳市朗驰欣创科技股份有限公司,广东深圳518000

出　　处：《红外与激光工程》2022年第8期287-295,共9页Infrared and Laser Engineering

基　　金：国家自然科学基金(11602292);四川省科技支撑计划(2021YFG0380)。

摘　　要：为解决低质量红外图像细节模糊、对比度低等问题,提出了并行多特征提取网络的红外图像增强方法,设计了结构特征映射网络和双尺度特征提取网络。结构特征映射网络用于建立全局结构特征权重,以保持原始图像的空间结构信息。双尺度特征提取网络采用多尺度卷积层和融合多空洞卷积的注意力,增强网络对上下文信息的关注力,提升网络对感兴趣区域的特征提取能力,同时学习不同尺度的特征信息,完成双尺度间信息的交换,生成目标增强映射,实现目标区域细节纹理自适应增强。实验证明,所提方法能有效提高对比度,避免过增强,丰富图像细节纹理,减少伪影和光晕现象,在BSD200数据集上的PSNR与SSIM较典型的传统方法和深度学习方法分别提升了约37.35%、2.1%与25.94%、3.15%,在真实红外数据集上分别提升了约30.62%、1.04%与24.83%、2.08%,且对不同对比度因子的低质量图像,文中方法也具有良好的增强效果。To solve the problems of fuzzy details and low contrast of low-quality infrared images,a parallel multifeature extraction network for infrared image enhancement is proposed,and a structural feature mapping network and a two-scale feature extraction network are designed.The structural feature mapping network is used to establish the global structural feature weight to maintain the spatial structure information of the original images.The two-scale feature extraction network using multiscale convolutional layers and the attention mechanism fused dilated convolutions is applied to enhance the attention on contextual information,improve the feature extraction capability for regions of interest,and simultaneously learn feature information of different scales,complete the exchange of information of the two scales,and then generate a target enhancement map to achieve adaptive enhancement of detailed texture of target areas.Experiments have proven that the proposed method can effectively improve contrast,avoid overenhancement,enrich image details and textures,and reduce artifacts and halos.Compared with typical traditional methods and deep learning methods,the PSNR and SSIM on the BSD200 dataset are increased by approximately 37.35%,2.1%and 25.94%,3.15%,and increased by approximately 30.62%,1.04%and 24.83%,2.08%on real infrared images.The proposed method also has good generalization performance on low-quality images with different contrast factors as well.

关 键 词：红外图像 图像增强 深度学习 空洞卷积 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]