This paper presents an algorithm for infrared and visible image fusion using significance detection and Convolutional Neural Networks with the aim of integrating discriminatory features and improving the overall quality of visual perception. Firstly, a global contrast-based significance detection algorithm is applied to the infrared image, so that salient features can be extracted, highlighting high brightness values and suppressing low brightness values and image noise. Secondly, a special loss function is designed for infrared images to guide the extraction and reconstruction of features in the network, based on the principle of salience detection, while the more mainstream gradient loss is used as the loss function for visible images in the network. Afterwards, a modified residual network is applied to complete the extraction of features and image reconstruction. Extensive qualitative and quantitative experiments have shown that fused images are sharper and contain more information about the scene, and the fused results look more like high-quality visible images. The generalization experiments also demonstrate that the proposed model has the ability to generalize well, independent of the limitations of the sensor. Overall, the algorithm proposed in this paper performs better compared to other state-of-the-art methods.

Color constancy methods are generally based on a simplifying assumption that the spectral distribution of a light source is uniform across scenes. However, in reality, this assumption is often violated because of the presence of multiple light sources, that is, more than two illuminations. In this paper, we propose a unique cascade network of deep multi-scale supervision and single-scale estimation (CN-DMS4) to estimate multi-illumination. The network parameters are supervised and learned from coarse to fine in the training process and estimate only the final thinnest level illumination map in the illumination estimation process. Furthermore, to reduce the influence of the color channel on the Euclidean distance or the pixel-level angle error, a new loss function with a channel penalty term is designed to optimize the network parameters. Extensive experiments are conducted on single and multi-illumination benchmark datasets. In comparison with previous multi-illumination estimation methods, our proposed method displays a partial improvement in terms of quantitative data and visual effect, which provides the future research direction in end-to-end multi-illumination estimation.