The images captured by a single sensor are often limited. How to use multi-sensor images has important value. For example, the imaging conditions of visible camera are relatively harsh, while the infrared camera can operate in all-day and all-weather and has longer visual distance. For better visual presentation and subsequent perception tasks, we focused on the infrared and visible image fusion based on auto-encoder. Specifically, we proposed a fusion strategy based on regional attention and a multi-scale convolution layer. The fusion strategy based on regional attention divides a image into several regions and adopts different fusion strategy for different regions. Multi-scale convolution layer is to capture the features of different receptive fields and improve the semantic representation ability of the encoder. From detailed experimental results, we can see that the optimized fusion algorithm is more robust, reduces the sensitivity to the classifier, and keeps more textures of background.
Infrared-visible image fusion based on regional attention auto-encoder
Third International Computing Imaging Conference (CITA 2023) ; 2023 ; Sydney, Australia
Proc. SPIE ; 12921
2023-11-06
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Infrared long-wave multispectral image reconstruction based on auto-encoder network
British Library Conference Proceedings | 2019
|A GAN-based visible and infrared image fusion algorithm
British Library Conference Proceedings | 2021
|BLURRED IMAGE REGION DETECTION BASED ON STACKED AUTO-ENCODER
British Library Conference Proceedings | 2018
|Deep Supervised Auto-encoder Hashing for Image Retrieval
British Library Conference Proceedings | 2018
|