Abstract: In this paper, we investigate the continual learning of Vision Transformers (ViT) for the challenging exemplar-free scenario, with special focus on how to efficiently distill the knowledge of its crucial self-attention mechanism (SAM). Our work takes an initial step towards a surgical investigation of SAM for designing coherent continual learning methods in ViTs. We first carry out an evaluation of established continual learning regularization techniques. We then examine the effect of regularization when applied to two key enablers of SAM: (a) the contextualized embedding layers, for their ability to capture well-scaled representations with respect to the values, and (b) the prescaled attention maps, for carrying value-independent global contextual information. We depict the perks of each distilling strategy on two image recognition benchmarks (CIFAR100 and ImageNet-32) – while (a) leads to a better overall accuracy, (b) helps enhance the rigidity by maintaining competitive performances. Furthermore, we identify the limitation imposed by the symmetric nature of regularization losses. To alleviate this, we propose an asymmetric variant and apply it to the pooled output distillation (POD) loss adapted for ViTs. Our experiments confirm that introducing asymmetry to POD boosts its plasticity while retaining stability across (a) and (b). Moreover, we acknowledge low forgetting measures for all the compared methods, indicating that ViTs might be naturally inclined continual learners. 1

Abstract: Brightness induction is the modulation of the perceived intensity of an area by the luminance of surrounding areas. Recent neurophysiological evidence suggests that brightness information might be explicitly represented in V1, in contrast to the more common assumption that the striate cortex is an area mostly responsive to sensory information. Here we investigate possible neural mechanisms that offer a plausible explanation for such phenomenon. To this end, a neurodynamical model which is based on neurophysiological evidence and focuses on the part of V1 responsible for contextual influences is presented. The proposed computational model successfully accounts for well known psychophysical effects for static contexts and also for brightness induction in dynamic contexts defined by modulating the luminance of surrounding areas. This work suggests that intra-cortical interactions in V1 could, at least partially, explain brightness induction effects and reveals how a common general architecture may account for several different fundamental processes, such as visual saliency and brightness induction, which emerge early in the visual processing pathway.

Abstract: Much of the neural machinery of the early visual cortex, from the extraction of local orientations to contextual modulations through lateral interactions, is thought to have developed to provide a sparse encoding of contour in natural scenes, allowing the brain to process efficiently most of the visual scenes we are exposed to. Certain visual stimuli, however, cause visual stress, a set of adverse effects ranging from simple discomfort to migraine attacks, and epileptic seizures in the extreme, all phenomena linked with an excessive metabolic demand. The theory of efficient coding suggests a link between excessive metabolic demand and images that deviate from natural statistics. Yet, the mechanisms linking energy demand and image spatial content in discomfort remain elusive. Here, we used theories of visual coding that link image spatial structure and brain activation to characterize the response to images observers reported as uncomfortable in a biologically based neurodynamic model of the early visual cortex that included excitatory and inhibitory layers to implement contextual influences. We found three clear markers of aversive images: a larger overall activation in the model, a less sparse response, and a more unbalanced distribution of activity across spatial orientations. When the ratio of excitation over inhibition was increased in the model, a phenomenon hypothesised to underlie interindividual differences in susceptibility to visual discomfort, the three markers of discomfort progressively shifted toward values typical of the response to uncomfortable stimuli. Overall, these findings propose a unifying mechanistic explanation for why there are differences between images and between observers, suggesting how visual input and idiosyncratic hyperexcitability give rise to abnormal brain responses that result in visual stress.

Abstract: A method for estimating the non-linear gamma transfer function of liquid–crystal displays (LCDs) without the need of a photometric measurement device was described by Xiao et al. (2011) [1]. It relies on observer’s judgments of visual luminance by presenting eight half-tone patterns with luminances from 1/9 to 8/9 of the maximum value of each colour channel. These half-tone patterns were distributed over the screen both over the vertical and horizontal viewing axes. We conducted a series of photometric and psychophysical measurements (consisting in the simultaneous presentation of half-tone patterns in each trial) to evaluate whether the angular dependency of the light generated by three different LCD technologies would bias the results of these gamma transfer function estimations. Our results show that there are significant differences between the gamma transfer functions measured and produced by observers at different viewing angles. We suggest appropriate modifications to the Xiao et al. paradigm to counterbalance these artefacts which also have the advantage of shortening the amount of time spent in collecting the psychophysical measurements.

Abstract: Generative Adversarial Networks (GANs) have recently demonstrated to successfully approximate complex data distributions. A relevant extension of this model is conditional GANs (cGANs), where the introduction of external information allows to determine specific representations of the generated images. In this work, we evaluate encoders to inverse the mapping of a cGAN, i.e., mapping a real image into a latent space and a conditional representation. This allows, for example, to reconstruct and modify real images of faces conditioning on arbitrary attributes.
Additionally, we evaluate the design of cGANs. The combination of an encoder
with a cGAN, which we call Invertible cGAN (IcGAN), enables to re-generate real
images with deterministic complex modifications.

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: Remote sensing applications, impacted by acquisition season and sensor variety, require high-resolution images. Transformer-based models improve satellite image super-resolution but are less effective than convolutional neural networks (CNNs) at extracting local details, crucial for image clarity. This paper introduces SWViT-RRDB, a new deep learning model for satellite imagery super-resolution. The SWViT-RRDB, combining transformer with convolution and attention blocks, overcomes the limitations of existing models by better representing small objects in satellite images. In this model, a pipeline of residual fusion group (RFG) blocks is used to combine the multi-headed self-attention (MSA) with residual in residual dense block (RRDB). This combines global and local image data for better super-resolution. Additionally, an overlapping cross-attention block (OCAB) is used to enhance fusion and allow interaction between neighboring pixels to maintain long-range pixel dependencies across the image. The SWViT-RRDB model and its larger variants outperform state-of-the-art (SoTA) models on two different satellite datasets in terms of PSNR and SSIM.

Abstract: Comic books constitute an important cultural heritage asset in many countries. Digitization combined with subsequent document understanding enable direct content-based search as opposed to metadata only search (e.g. album title or author name). Few studies have been done in this direction. In this work we detail a novel approach for the automatic text localization in scanned comics book pages, an essential step towards a fully automatic comics book understanding. We focus on speech text as it is semantically important and represents the majority of the text present in comics. The approach is compared with existing methods of text localization found in the literature and results are presented.