A Convolutional Neural Network for Learning Local Feature Descriptors on Multispectral Images

This work presents a novel convolutional neural network, termed multispectral features network (MF-Net), for learning local feature descriptors in multispectral images. Unlike most existing solutions, which primarily handle images from the visible light spectrum, we propose a learning-based method that deals with image data acquired from different spectrum bands. To design our convolutional neural network, we introduce a new layer that incorporates Log-Gabor filters to enhance the network capability to work with nonlinear intensity variations in images captured from different electromagnetic frequencies spectrum. This layer, entitled mapping layer, can be easily integrated into different network architectures. To demonstrate the efficacy and limitations of our method, we went on experiments with two distinct datasets extensively used in previous works composed of image pairs from the visible spectrum and the infrared spectrum. Experimental results with datasets containing images obtained from visible light and infrared spectrum show that our method can accurately match features, outperforming some state-of-the-art learning-based algorithms.