Lightweight Infrared and Visible Image Fusion Network with Depthwise Separable Convolution

This paper presents a lightweight infrared-visible image fusion network based on depthwise separable convolutions, designed to address the issues of high computational complexity, substantial computational workload, and inadequate feature extraction capabilities inherent in traditional image fusion methods. The framework utilizes a dual-encoder configuration to independently derive features from infrared and visible images, employing depthwise separable convolutions to minimize model parameters and boost computational efficiency. Furthermore, by incorporating SSIM and patchNCE as loss functions, the framework seeks to retain more intricate details from the source images and elevate the quality of the fused images. Experimental findings reveal that, when compared to seven other established methods on two publicly accessible datasets, the proposed method achieves an 3.6% improvement in the EN metric, a 8.9% enhancement in the SF metric, a 8.7% increase in the AG metric, a 9.7% rise in the SD metric, and a 42.86% acceleration in processing speed.