CTFusion: CNN-transformer-based self-supervised learning for infrared and visible image fusion

Abstract

<p>Infrared and visible image fusion (IVIF) is devoted to extracting and integrating useful complementary information from muti-modal source images. Current fusion methods usually require a large number of paired images to train the models in supervised or unsupervised way. In this paper, we propose CTFusion, a convolutional neural network (CNN)-Transformer-based IVIF framework that uses self-supervised learning. The whole framework is based on an encoder-decoder network, where encoders are endowed with strong local and global dependency modeling ability via the CNN-Transformer-based feature extraction (CTFE) module design. Thanks to the development of self-supervised learning, the model training does not require ground truth fusion images with simple pretext task. We designed a mask reconstruction task according to the characteristics of IVIF, through which the network can learn the characteristics of both infrared and visible images and extract more generalized features. We evaluated our method and compared it to five competitive traditional and deep learning-based methods on three IVIF benchmark datasets. Extensive experimental results demonstrate that our CTFusion can achieve the best performance compared to the state-of-the-art methods in both subjective and objective evaluations.</p>