SuPeR: Milano Retinex implementation exploiting a regular image grid

Improving the perception of low-light enhanced images

Improving images captured under low-light conditions has become an important topic in computational color imaging, as it has a wide range of applications. Most current methods are either based on handcrafted features or on end-to-end training of deep neural networks that mostly focus on minimizing some distortion metric -such as PSNR or SSIM- on a set of training images. However, the minimization of distortion metrics does not mean that the results are optimal in terms of perception (i.e. perceptual quality). As an example, the perception-distortion trade-off states that, close to the optimal results, improving distortion results in worsening perception. This means that current low-light image enhancement methods -that focus on distortion minimization- cannot be optimal in the sense of obtaining a good image in terms of perception errors. In this paper, we propose a post-processing approach in which, given the original low-light image and the result of a specific method, we are able to obtain a result that resembles as much as possible that of the original method, but, at the same time, giving an improvement in the perception of the final image. More in detail, our method follows the hypothesis that in order to minimally modify the perception of an input image, any modification should be a combination of a local change in the shading across a scene and a global change in illumination color. We demonstrate the ability of our method quantitatively using perceptual blind image metrics such as BRISQUE, NIQE, or UNIQUE, and through user preference tests.

Mathematical insights into the original Retinex algorithm for image enhancement

The Retinex theory, originally developed by Land and McCann as a computation model of the human color sensation, has become, with time, a pillar of digital image enhancement. In this area, the Retinex algorithm is widely used to improve the quality of any input image by increasing the visibility of its content and details, enhancing its colorfulness, and weakening, or even removing, some undesired effects of the illumination. The algorithm was originally described by its creators in terms of a sequence of image processing operations and was not fully formalized mathematically. Later, works focusing on aspects of the original formulation and adopting some of its principles tried to frame the algorithm within a mathematical formalism: this yielded every time a partial rendering of the model and resulted in several interesting model variants. The purpose of the present work is to fill a gap in the Retinex-related literature by providing a complete mathematical formalization of the original Retinex algorithm. The overarching goals of this work are to provide mathematical insights into the Retinex theory, promote awareness of the use of the model within image enhancement, and enable better appreciation of differences and similarities with later models based on Retinex principles. For this purpose, we compare our model with others proposed in the literature, paying particular attention to the work published in 2005 by Provenzi and others.

Performance comparison of image enhancers with and without deep learning

Image enhancement is a computational procedure to improve visibility of details and content of an input image. Several image enhancement algorithms have been developed thus far, from traditional methods that process a single image based on physical models of image acquisition and formation to recent deep learning techniques, where enhancement models are learned from data. Here, we empirically compare a set of traditional and deep learning enhancers, which we selected as representing different methodologies for the improvement of poorly illuminated images. Our experiments are conducted on public data and show that, although all the considered enhancers improve the visibility of the image content and details, the deep-learning-based methods generally produce less noisy images than the traditional ones. This last outcome has to be carefully considered when enhancers are used as preprocessing for algorithms that are sensitive to noise. As a case study, and with the purpose to promote more aware usage of these two groups of enhancers in computer vision applications, we discuss the impact of image enhancement in the framework of image retrieval performed through two popular algorithms, i.e., SIFT and ORB, implementing different image descriptions and having different sensitivities to noise.

An Image Contrast Measure Based on Retinex Principles

The image contrast is a feature capturing the variation of the image signal across the space. Such a feature is very useful to describe the local image structure at different scales and thus it is relevant to many computer vision applications, like image/texture retrieval and object recognition. In this work, we present MiRCo, a novel measure of image contrast derived from the Retinex theory. MiRCo is robust against in-plane rotations and light changes at multiple scales. Thanks to these properties, MiRCo enables an accurate and robust description of the local image structure. Here we describe and discuss the mathematical insights of MiRCo also in comparison with other popular contrast measures.

Generalized equation for real-world image enhancement by Milano Retinex family

Milano Retinexes are spatial color algorithms grounded on the Retinex theory and widely applied to enhance the visual content of real-world color images. In this framework, they process the color channels of the input image independently and re-scale channel by channel the intensity of each pixel p by the so-called local reference white, i.e., a strictly positive value, computed by reworking a set of features sampled around p. The neighborhood of p to be sampled, its sampling, the features to be processed, as well as the mathematical model for the computation of the local reference white vary from algorithm to algorithm, determining different levels of enhancement. Based on the analysis of a group of Milano Retinexes, this work proves that the Milano Retinex local reference whites can be expressed by a generalized equation whose parameters model specific aspects of the Milano Retinex spatial color processing. In particular, tuning these parameters leads to different Milano Retinex implementations. This study contributes to a better understanding of the similarities and differences among the members of the Milano Retinex family, and provides new taxonomic schemes of them based on their own mathematical properties.