Federated Learning (FL) enables decentralized training across multiple clients without exposing local data. However, standard FL algorithms struggle under heterogeneous data distributions, often leading to global models that fail to generalize and personalized models that ignore collaborative benefits. Existing approaches based on model interpolation, such as APFL, offer a promising direction but lack mechanisms to adaptively select which model to use at inference time. In this work, we propose FLProtector, a dual-model framework in which each client learns a personalized increment over a shared global model, and dynamically selects between the two during inference based on novelty detection. This decision is guided by a client-specific autoencoder trained on local data to identify out-of-distribution inputs. Our method also incorporates a robust aggregation strategy based on gradient consistency, reducing the impact of clients whose updates deviate from the global optimization path. Experiments on the Digit-five benchmark under both fully and partially heterogeneous scenarios demonstrate that FLProtector outperforms classical and state-of-the-art personalized FL baselines, offering superior personalization without compromising global generalization. Notably, it achieves strong results without requiring sensitive hyperparameter tuning.

A benchmark of saliency models performance with a synthetic image dataset is provided. Model performance is evaluated through saliency metrics as well as the influence of model inspiration and consistency with human psychophysics. SID4VAM is composed of 230 synthetic images, with known salient regions. Images were generated with 15 distinct types of low-level features (e.g. orientation, brightness, color, size...) with a target-distractor pop-out type of synthetic patterns. We have used Free-Viewing and Visual Search task instructions and 7 feature contrasts for each feature category. Our study reveals that state-of-the-art Deep Learning saliency models do not perform well with synthetic pattern images, instead, models with Spectral/Fourier inspiration outperform others in saliency metrics and are more consistent with human psychophysical experimentation. This study proposes a new way to evaluate saliency models in the forthcoming literature, accounting for synthetic images with uniquely low-level feature contexts, distinct from previous eye tracking image datasets.

General dynamic scenes involve multiple rigid and flexible objects, with relative and common motion, camera induced or not. The complexity of the motion events together with their strong spatio-temporal correlations make the estimation of dynamic visual saliency a big computational challenge. In this work, we propose a computational model of saliency based on the assumption that perceptual relevant information is carried by high-order statistical structures. Through whitening, we completely remove the second-order information (correlations and variances) of the data, gaining access to the relevant information. The proposed approach is an analytically tractable and computationally simple framework which we call Dynamic Adaptive Whitening Saliency (AWS-D). For model assessment, the provided saliency maps were used to predict the fixations of human observers over six public video datasets, and also to reproduce the human behavior under certain psychophysical experiments (dynamic pop-out). The results demonstrate that AWS-D beats state-of-the-art dynamic saliency models, and suggest that the model might contain the basis to understand the key mechanisms of visual saliency. Experimental evaluation was performed using an extension to video of the well-known methodology for static images, together with a bootstrap permutation test (random label hypothesis) which yields additional information about temporal evolution of the metrics statistical significance.

It is of great benefit to have advance knowledge of human visual target acquisition performance for targets or other relevant objects. However, search performance inherently shows a large variance and depends strongly on prior knowledge of the perceived scene. A typical search experiment therefore requires a large number of observers to obtain statistically reliable data. Moreover, measuring target acquisition performance in field situations is usually impractical and often very costly or even dangerous. The paper presents a method for characterizing information of a target relative to its background. The resultant computational measures are then applied to quantify the visual distinctness of targets in complex natural backgrounds from digital imagery. A generalization of the Kullback-Leibler joint information gain of various random variables is shown to correlate strongly with visual target distinctness as estimated by human observers. Bootstrap methods for assessing statistical accuracy were used to produce this inference.

We show that a notion of congruence in statistical structure across 2D frequency bands produces a useful definition of visual patterns for perceiving target distinctness. In order to reach such a conclusion, firstly, the notion of congruence is used to induce a representational model for 2D images. Secondly, the visual-pattern based representational model is used to define a visual target distinctness metric that involves applying a simple decision rule over the distances between the visual patterns. Finally, a relation is established between the computational distinctness metric and psychophysical target distinctness estimates.

To discover clustered microcalcifications in computerized mammogram image analysis firstly we must define what we mean by calcifications in digitized mammographic images. Here we show that given an objective function based on the invariance across logGabor filters actively responding to different aspects of obvious and subtle objects in digitized mammograms, the optimal notion of microcalcification can be defined as congruence in local entropy at attentional points across a range of 2D frequency bands. We reach this conclusion from several experiments in which the notion of microcalcification as congruence in statistical structure across frequency bands was applied on a number of digitized mammograms from the MIAS and Nijmegen databases in order to investigate the segregation of obvious and subtle objects.

This paper describes a system for the automatically learned partitioning of visual patterns in 2D images, based on sophisticated band-pass filtering with fixed scale and orientation sensitivity. The visual patterns are defined as the features which have the highest degree of alignment in the statistical structure across different frequency bands. The analysis reorganizes the image according to an invariance constraint in statistical structure and consists of three stages: pre-attentive stage, integration stage, and learning stage. The first stage takes the input image and performs filtering with log-Gabor filters. Based on their responses, activated filters which are selectively sensitive to patterns in the image are short listed. In the integration stage, common grounds between several activated sensors are explored. The filtered responses are analyzed through a family of statistics. For any given two activated filters, a distance between them is derived via distances between their statistics. The third stage performs cluster partitioning for learning the subspace of log-Gabor filters needed to partition the image data. The clustering is based on a dissimilarity measure intended to highlight scale and orientation invariance of the responses. The technique is illustrated on real and simulated data sets. Finally, this paper presents a computational visual distinctness measure computed from the image representational model based on visual patterns. Experiments are performed to investigate its relation to distinctness as measured by human observers.

This paper analyzes how the natural scales of the shapes in 2D images can be extracted. Spatial information is analyzed by multiple units sensitive to both spatial and spatial-frequency variables. Scale estimates of the relevant shapes are constructed only from strongly responding detectors. The meaningful structures in the response of a detector (computed through 2D Gabor filtering) are, at their natural level of resolution, relatively sharp and have well-defined boundaries. A natural scale is so defined as a level producing local minimum of a function that returns the relative sharpness of the detector response filtered over a range of scales. In a second stage, to improve a first crude estimate of the local scale, the criterion is also rewritten to directly select scales at locations of significant features of each activated detector.