基于密集连接的红外可见光图像融合方法  

Infrared and visible light image fusion method based on dense connection

作　　者：李嘉元 程江华[1] 刘通[1] 程榜 潘乐昊 LI Jiayuan;CHENG Jianghua;LIU Tong;CHENG Bang;PAN Lehao(School of Electronic Science,National University of Defense Technology,Changsha Hunan 410073,China)

机构地区：[1]国防科技大学电子科学学院,长沙410073

出　　处：《计算机应用》2023年第S02期163-167,共5页journal of Computer Applications

基　　金：湖南省自然科学基金资助项目(2020JJ4670)。

摘　　要：针对基于人工智能的红外可见光图像的融合方法中神经网络网络设计复杂、参数量和推理计算量较大的问题,提出一种轻量化网络结构。首先,使用两个相同的密集连接块提取图像特征;然后,对红外图像和可见光图像使用共同的卷积核,以减少参数量;最后,利用残差连接保留梯度信息,同时基于像素显著性原则设计网络损失函数。在TNO和M3FD数据集上的实验结果表明,与U2Fusion经典算法相比,所提方法的参数量和推理计算量分别降低了83.5%和87.5%,并保持了良好的融合效果。A light-weight neural network structure was proposed to address the issues of complex neural network design,numerous parameters and large inference computation in the image fusion method based on artificial intelligence.Firstly,two identical dense connection blocks were used to extract image features.Then,the common convolution kernels were used for infrared images and visible images to reduce the number of parameters.Finally,the residual connection was used to retain the gradient information,and the network loss function was designed based on the principle of pixel saliency.The experimental results on TNO and M3FD datasets demonstrate that compared with the U2Fusion,the number of parameters and amount of inference computation reduced by 83.5%and 87.5%,while maintaining good fusion performance.

关 键 词：图像融合 轻量化网络 密集连接 深度学习 红外图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]