DePF: A Novel Fusion Approach Based on Decomposition Pooling for Infrared and Visible Images
Hui Li, Yongbiao Xiao, Chunyang Cheng, Zhongwei Shen, Xiaoning Song
Abstract
Infrared and visible image fusion is a crucial technique in the field of computer vision, aiming to create synthetic images that simultaneously capture salient features and rich texture details. These fused images play a pivotal role in enhancing various downstream tasks. However, existing fusion methods often encounter challenges such as texture loss and deficiencies in edge information, resulting in less than optimal fusion outcomes. Additionally, straight-forward up-sampling techniques struggle to preserve source information from multi-scale features. To tackle these issues, a novel fusion network based on the decomposition pooling (de-pooling) manner is proposed, termed as DePF. DePF features a de-pooling-based encoder designed to extract multi-scale image and detail features from source images concurrently. Furthermore, a spatial attention model aggregates these salient features, and the decoder employs a de-pooling reversed operation instead of the typical up-sampling operator to reconstruct the fused features. Unlike conventional max-pooling techniques, the de-pooling layer preserves abundant detail information, facilitating a richer texture and multi-scale information retention during the reconstruction phase. Significantly, our approach exhibits remarkable efficiency, requiring merely 23 ms to integrate a pair of infrared and visible images, each with dimensions of 640 × 480. Furthermore, empirical findings corroborate the exceptional fusion efficacy of our methodology in the domains of object detection and noise-related assessments, surpassing the performance of contemporary techniques within numerous image fusion benchmarks.