Abstract: Convolutional Neural Networks (CNNs) trained for object recognition tasks present representational capabilities approaching to primate visual systems [1]. This provides a computational framework to explore how image features
are efficiently represented. Here, we dissect a trained CNN
[2] to study how color is represented. We use a classical methodology used in physiology that is measuring index of selectivity of individual neurons to specific features. We use ImageNet Dataset [20] images and synthetic versions
of them to quantify color tuning properties of artificial neurons to provide a classification of the network population.
We conclude three main levels of color representation showing some parallelisms with biological visual systems: (a) a decomposition in a circular hue space to represent single color regions with a wider hue sampling beyond the first
layer (V2), (b) the emergence of opponent low-dimensional spaces in early stages to represent color edges (V1); and (c) a strong entanglement between color and shape patterns representing object-parts (e.g. wheel of a car), objectshapes (e.g. faces) or object-surrounds configurations (e.g. blue sky surrounding an object) in deeper layers (V4 or IT).

Abstract: Extraction of spatio-chromatic features from color images is usually performed independently on each color channel. Usual 3D color spaces, such as RGB, present a high inter-channel correlation for natural images. This correlation can be reduced using color-opponent representations, but the spatial structure of regions with small color differences is not fully captured in two generic Red-Green and Blue-Yellow channels. To overcome these problems, we propose a new color coding that is adapted to the specific content of each image. Our proposal is based on two steps: (a) setting the number of channels to the number of distinctive colors we find in each image (avoiding the problem of channel correlation), and (b) building a channel representation that maximizes contrast differences within each color channel (avoiding the problem of low local contrast). We call this approach more-than-three color coding (MTT) to enhance the fact that the number of channels is adapted to the image content. The higher color complexity an image has, the more channels can be used to represent it. Here we select distinctive colors as the most predominant in the image, which we call color pivots, and we build the new color coding using these color pivots as a basis. To evaluate the proposed approach we measure its efficiency in an image categorization task. We show how a generic descriptor improves its performance at the description level when applied on the MTT coding.

Abstract: Visual recognition is one of the main problems in computer vision that attempts to solve image understanding by deciding what objects are in images. This problem can be computationally solved by using relevant sets of visual features, such as edges, corners, color or more complex object parts. This thesis contributes to how color features have to be represented for recognition tasks.
Image features can be extracted following two different approaches. A first approach is defining handcrafted descriptors of images which is then followed by a learning scheme to classify the content (named flat schemes in Kruger et al. (2013). In this approach, perceptual considerations are habitually used to define efficient color features. Here we propose a new flat color descriptor based on the extension of color channels to boost the representation of spatio-chromatic contrast that surpasses state-of-the-art approaches. However, flat schemes present a lack of generality far away from the capabilities of biological systems. A second approach proposes evolving these flat schemes into a hierarchical process, like in the visual cortex. This includes an automatic process to learn optimal features. These deep schemes, and more specifically Convolutional Neural Networks (CNNs), have shown an impressive performance to solve various vision problems. However, there is a lack of understanding about the internal representation obtained, as a result of automatic learning. In this thesis we propose a new methodology to explore the internal representation of trained CNNs by defining the Neuron Feature as a visualization of the intrinsic features encoded in each individual neuron. Additionally, and inspired by physiological techniques, we propose to compute different neuron selectivity indexes (e.g., color, class, orientation or symmetry, amongst others) to label and classify the full CNN neuron population to understand learned representations.
Finally, using the proposed methodology, we show an in-depth study on how color is represented on a specific CNN, trained for object recognition, that competes with primate representational abilities (Cadieu et al (2014)). We found several parallelisms with biological visual systems: (a) a significant number of color selectivity neurons throughout all the layers; (b) an opponent and low frequency representation of color oriented edges and a higher sampling of frequency selectivity in brightness than in color in 1st layer like in V1; (c) a higher sampling of color hue in the second layer aligned to observed hue maps in V2; (d) a strong color and shape entanglement in all layers from basic features in shallower layers (V1 and V2) to object and background shapes in deeper layers (V4 and IT); and (e) a strong correlation between neuron color selectivities and color dataset bias.

Abstract: We present a method to estimate the direction and color of the scene light source from a single image. Our method is based on two main ideas: (a) we use a new synthetic dataset with strong shadow effects with similar constraints to the SID dataset; (b) we define a deep architecture trained on the mentioned dataset to estimate the direction and color of the scene light source. Apart from showing good performance on synthetic images, we additionally propose a preliminary procedure to obtain light positions of the Multi-Illumination dataset, and, in this way, we also prove that our trained model achieves good performance when it is applied to real scenes.

Abstract: In this thesis, we explore how to estimate the effects of the light interacting with the scene objects from a single image. To achieve this goal, we focus on recovering intrinsic components like reflectance, shading, or light properties such as color and position using deep architectures. The success of these approaches relies on training on large and diversified image datasets. Therefore, we present several contributions on this such as: (a) a data-augmentation technique; (b) a ground-truth for an existing multi-illuminant dataset; (c) a family of synthetic datasets, SID for Surreal Intrinsic Datasets, with diversified backgrounds and coherent light conditions; and (d) a practical pipeline to create hybrid ground-truths to overcome the complexity of acquiring realistic light conditions in a massive way. In parallel with the creation of datasets, we trained different flexible encoder-decoder deep architectures incorporating physical constraints from the image formation models.
In the last part of the thesis, we apply all the previous experience to two different problems. Firstly, we create a large hybrid Doc3DShade dataset with real shading and synthetic reflectance under complex illumination conditions, that is used to train a two-stage architecture that improves the character recognition task in complex lighting conditions of unwrapped documents. Secondly, we tackle the problem of single image scene relighting by extending both, the SID dataset to present stronger shading and shadows effects, and the deep architectures to use intrinsic components to estimate new relit images.

Abstract: the performance of colour correction algorithms are dependent on the reflectance sets used. Sometimes, when the testing reflectance set is changed the ranking of colour correction algorithms also changes. To remove dependence on dataset we can
make assumptions about the set of all possible reflectances. In the Maximum Ignorance with Positivity (MIP) assumption we assume that all reflectances with per wavelength values between 0 and 1 are equally likely. A weakness in the MIP is that it fails to take into account the correlation of reflectance functions between
wavelengths (many of the assumed reflectances are, in reality, not possible).
In this paper, we take the view that the maximum ignorance assumption has merit but, hitherto it has been calculated with respect to the wrong coordinate basis. Here, we propose the Discrete Cosine Maximum Ignorance assumption (DCMI), where
all reflectances that have coordinates between max and min bounds in the Discrete Cosine Basis coordinate system are equally likely.
Here, the correlation between wavelengths is encoded and this results in the set of all plausible reflectances ’looking like’ typical reflectances that occur in nature. This said the DCMI model is also a superset of all measured reflectance sets.
Experiments show that, in colour correction, adopting the DCMI results in similar colour correction performance as using a particular reflectance set.

Abstract: This paper presents a first approach to build colour and structural descriptors for information retrieval on a people database. Queries are formulated in terms of their appearance that allows to seek people wearing specific clothes of a given colour name or texture. Descriptors are automatically computed by following three essential steps. A colour naming labelling from pixel properties. A region seg- mentation step based on colour properties of pixels combined with edge information. And a high level step that models the region arrangements in order to build clothes structure. Results are tested on large set of images from real scenes taken at the entrance desk of a building