LMDFusion: A lightweight infrared and visible image fusion network for substation equipment based on mask and residual dense connection

Abstract

Lightweight infrared and visible image fusion is a challenge problem for image-based fault detection. In this paper, a lightweight infrared and visible image fusion network is proposed based on mask and residual dense connection called LMDFusion. In this method, the loss function is designed by utilizing image masks to assist the training of the network, and the residual dense connection is used to better preserve the deep features of each convolutional layer to make the generated fusion image well retain the thermal radiation target and background texture structure from source image. Additionally, depthwise separable convolution is used to replace standard convolution to make our model more lightweight and suitable for deployment on mobile compared to most advanced fusion models. Although the purpose of our research is mainly for image fusion of electrical substation equipment, it still performs well in other public infrared and visible datasets. Experimental results on different datasets indicate that the proposed method exhibits outstanding performance in terms of subjective visual evaluation and objective evaluation metrics such as entropy (EN), standard deviation (SD), mutual information (MI) and pixel-based visual information fidelity (VIFP).