International audience

Skin cancer is among the most prevalent and deadly types of cancer. Dermatologists mostly use visual cues to diagnose this illness. The classification of multiclass skin cancer is challenging due to the fine-grained variability in the appearance of its several diagnostic categories. Advances in deep neural networks have led to a significant increase in skin lesion classification methods in recent years, with more performante models like vision transformers (ViTs) emerging and improving skin lesion classification performance. Our explainable skin cancer classification with ViT and Grad-CAM (XSC-ViT) is presented in this paper. This uses a specifically designed ViT to classify skin cancer. Our solution integrates L2 regularization in some ViT layers to prevent overfitting and applies the Grad-CAM (Gradient-weighted Class Activation Mapping) method for decision explainability. Our approach demonstrated remarkable performance results for skin cancer classification with an accuracy of 92%, precision of 91.73%, recall of 91.46%, F1 score of 91.51%, and loss of 22.05% when tested on the HAM10000 dataset.

Handwriting analysis has proven to be a valuable tool in diagnosing neurological diseases and neurodevelopmental disorders. Among these, Attention Deficit Hyperactivity Disorder (ADHD) stands out as a significant focus. ADHD affects how the brain manages attention, impulsivity, and hyperactivity. In this context, this study proposes a novel method to ADHD prediction in children by integrating handwriting-based features within a hybrid learning analysis. Specifically, we focus on combining the strengths of multiple base models in a voting classification framework to achieve improved accuracy and robustness. Such a proposed hybrid learning approach is compared with single learning as well as ensemble learning approaches. In addition, we focus on feature engineering that explores handwriting patterns and dynamics to identify the most informative characteristics for accurate ADHD detection. The experimental results demonstrate the effectiveness of the proposed method, showing promising accuracy in distinguishing children with ADHD from healthy subjects.

International audience

Deep Neural Networks achieve outstanding results; however, their reliance on a static environment with fixed data poses challenges in dynamic scenarios where data continuously evolves. Being capable of learning, adapting, and generalizing continually in a scalable, successful, and efficient manner is crucial for the sustainable development of AI systems. The classical solution of retraining the model using both old and new data is time-consuming and expensive. Continual Learning tackles the problem of learning new data distributions without the need for retraining from scratch. Furthermore, the task of recognizing unlabeled images using previously acquired knowledge becomes challenging, particularly when the new data needs to be incrementally annotated without starting the training process from scratch. To gain a deeper understanding of how “Black Box” neural networks make decisions, it is important to visualize components inside the model that affect the error rate throughout the decision-making process. The Continual Self-Learning model on label-less historical digits yields increasingly perceptive interpretations. This paper aims to establish a literature review of the latest advances in continual learning for computer vision tasks , to articulate catastrophic forgetting using Explainable Artificial Intelligence on both split MNIST and the historical digit dataset DIDA, and to shed light on important but still understudied topics.

Self-learning has demonstrated remarkable performance in computer vision tasks. However, dealing with unlabeled data that needs to be recognized based on previously acquired knowledge is crucial. Specifically, in Continual Learning where new data is incrementally learned, without the need for training from scratch, models often suffer from Catastrophic forgetting. In this study, we investigate the potential of continual self-learning, using LeNet architecture, to address the issue of learning and recognizing unlabeled historical digits by transferring the knowledge gained from the Split-MNIST dataset. The application of Continual self-learning to unlabeled historical digits unlocks the potential for machines to gain a deeper understanding of our digit-based history and provide increasingly insightful interpretations. To the best of our knowledge, this is the first work that investigates Continual Learning in the context of historical handwritten text recognition.

Telemedicine, particularly when conducted through unprotected connections, necessitates secure delivery of medical data. In the context of medical image transmission, watermarking techniques are frequently employed to enhance the security and privacy of digital content. This paper proposes a secure and blind watermarking algorithm based on spatial domain Least Significant Bit (LSB), chaotic sequences and key point detection using BRISK (Binary Robust Invariant Scalable Keypoints) and GFTT (Good Feature To Track) algorithms, combined with K-means clustering. Chaotic sequences are used for watermark encryption and embedding, ensuring robust security. The algorithm converts the message to binary format, then divides the cover medical image into three bands (red, green, blue) and hides the encrypted watermark within the green and blue bands. The merged bands (red, modified green, modified blue) create the final image with hidden watermark. The proposed approach has undergone experimental investigation within the field of fundus images, which are a specialized form of medical images capturing the back portion of the eye, known as the fundus. In this particular application, fundus images incorporate a watermark encrypting patient-related data and ensuring its integrity during transmission and storage. In addition, integrity check data and other important data are delivered to the recipient to verify the watermark. Through the experimental study, the proposed approach showcases its effectiveness and ability to withstand different attacks, making it highly suitable for secure telemedicine applications.

Federated Learning can ensure privacy by design through its unique approach to data analysis and model training. Adapting Machine Learning models into federated architectures is particularly interesting, especially in the context of anomaly detection and data preservation in time series data of smart meters. In this study, we explore Machine Learning models deployed and depicted on an open dataset to demonstrate the efficiency of Machine Learning based Federated Learning frameworks in preserving privacy. Through comprehensive experiments and evaluations, we highlight the significant privacy achieved through decentralized data, local model training and aggregated model updates. By harnessing the power of the federated learning, we offer a promising solution for preserving privacy in anomaly detection while enabling effective analysis of smart meter data.

Over the past few decades, handwriting recognition has made tremendous progress based on deep learning architectures. Most commonly, static learning has achieved impressive results on various datasets. However, this approach relies predominantly on fixed datasets and stationary environments, making it challenging to handle continuous streams of data without forgetting previously learned knowledge. Conversely, Continual learning offers a more efficient and flexible framework for learning from a stream of data over time. However, it can be challenging to compare the performance of these approaches due to differences in evaluation protocols and a lack of a common framework. Moreover, prior research has primarily focused on updating models to perform well on specific task or domain scenarios. Yet, the Class-Incremental Learning scenario, which involves learning new classes incrementally over time, has not been as well addressed as it is considered a more challenging task. To address this gap, this paper provides a comprehensive empirical comparison of some of the currently used continual learning strategies, specifically focusing on the case where the model must learn to recognize continuously new classes while the task-ID is unknown. We conducted a deep overview of recent research from different perspectives and show that well-known approaches such as EWC, LWF, and SI are not ideal for the class-IL scenarios. We discuss the results in more detail and highlight some of the limitations that we believe require further investigation.

Super‐resolution (SR) task has become an important research area due to the rapidly growing interest for high quality images in various computer vision and pattern recognition applications. This has led to the emergence of various SR approaches. According to the number of input images, two kinds of approaches could be distinguished: single or multi‐input based approaches. Certainly, processing multiple inputs could lead to an interesting output, but this is not the case mainly for textual image processing. This study focuses on single image‐based approaches. Most of the existing methods have been successfully applied on natural images. Nevertheless, their direct application on textual images is not enough efficient due to the specificities that distinguish these particular images from natural images. Therefore, SR approaches especially suited for textual images are proposed in the literature. Previous overviews of SR methods have been concentrated on natural images application with no real application on the textual ones. Thus, this study aims to tackle this lack by surveying methods that are mainly designed for enhancing low‐resolution textual images. The authors further criticise these methods and discuss areas which promise improvements in such task. To the best of the authors’ knowledge, this survey is the first investigation in the literature.

Current issues on textual image magnification have been focused on noise-free low-resolution images. Nevertheless, real circumstances are far from these assumptions and existing systems are generally confronted with noisy images; limiting thus the efficiency of the magnification process. The scope of this study is to propose a joint denoising and magnification system based on sparse coding to tackle such a problem. The underlying idea suggests the representation of an image patch by a linear combination of few elements from a suitable dictionary. The proposed system uses both online and offline learned dictionaries that are selected adaptively for each image patch of the input Low-Resolution (LR) noisy image to generate its corresponding noise-free High-Resolution (HR) version. In fact, the online learned dictionaries are trained on a clustered dataset of the image patches selected from the input image and used for the denoising purpose in order to take benefit of the non-local self-similarity assumption in textual images. For the offline learned dictionaries, they are trained on an external LR/HR image patch pair dataset and employed for the magnification purpose. The performance of the proposed system is evaluated visually and quantitatively on different LR noisy textual images and promising results are achieved when compared with other existing systems and conventional approaches dealing with such kind of images.

Sparse coding has shown to be an effective technique in solving various reconstruction tasks such as denoising, in painting, and resolution enhancement of natural images. In this paper, we explore the use of this technique specifically to deal with low-resolution and degraded textual images. Firstly, we propose a sparse coding based resolution enhancement approach to recover a textual image with higher resolution than the input low-resolution one. It is based on the use of multiple coupled dictionaries which are learned from a clustered training low-resolution/high-resolution patch-pair database. A reconstruction scheme is then suggested in order to adaptively select the appropriate dictionaries that are useful for better recovering each local patch. This approach can be applied for the magnification of both printed and handwritten characters. Secondly, we propose to integrate the magnification in a restoration framework specifically to denoise and reconstruct at the same time degraded characters. The performances of these propositions are evaluated on various types of degraded printed and handwritten textual images where loss of details and background noise exist. Promising results are achieved when compared with results of other existing approaches.

Sparse coding is widely known as a methodology where an input signal can be sparsely represented from a suitable dictionary. It was successfully applied on a wide range of applications like the textual image Super-Resolution. Nevertheless, its complexity limits enormously its application. Looking for a reduced computational complexity, a coupled dictionary learning approach is proposed to generate dual dictionaries representing coupled feature spaces. Under this approach, we optimize the training of a first dictionary for the high-resolution image space and then a second dictionary is simply deduced from the latter for the low-resolution image space. In contrast with the classical dictionary learning approaches, the proposed approach allows a noticeable speedup and a major simplification of the coupled dictionary learning phase both in terms of algorithm architecture and computational complexity. Furthermore, the resolution enhancement results achieved by applying the proposed approach on poorly resolved textual images lead to image quality improvements.

Global Mean Shift algorithm is an unsupervised clustering technique already applied for color document image segmentation. Nevertheless, its important computational cost limits its application for document images. The complexity of the global approach is explained by the intensive search of colors samples in the Parzen window to compute the vector oriented toward the mean. For making it more flexible, several attempts have tried to decrease the algorithm complexity mainly by adding spatial information or by reducing the number of colors to shift or even by selecting a reduced number of colors to estimate the means of density function. This paper presents a fast optimized Mean Shift with a much reduced computational cost. This algorithm uses both the discretisation of the shift and the integral image which allow the computation of means into the Parzen windows with a reduced and fixed number of operations. With the discretisation of the color space, the fast optimised MeanShift also memorizes all existing paths to avoid shifting again colors along similar path. Despite the square shape of the Parzen windows and the uniform kernel used, the results are very similar to those obtained by the global Mean Shift algorithm. The proposed algorithm is compared to the different existing implementation of similar algorithms found in the literature.

This paper addresses the problem of generating a super-resolved version of a low-resolution textual image by using Sparse Coding (SC) which suggests that image patches can be sparsely represented from a suitable dictionary. In order to enhance the learning performance and improve the reconstruction ability, we propose in this paper a multiple learned dictionaries based clustered SC approach for single text image super resolution. For instance, a large High-Resolution/Low-Resolution (HR/LR) patch pair database is collected from a set of high quality character images and then partitioned into several clusters by performing an intelligent clustering algorithm. Two coupled HR/LR dictionaries are learned from each cluster. Based on SC principle, local patch of a LR image is represented from each LR dictionary generating multiple sparse representations of the same patch. The representation that minimizes the reconstruction error is retained and applied to generate a local HR patch from the corresponding HR dictionary. The performance of the proposed approach is evaluated and compared visually and quantitatively to other existing methods applied to text images. In addition, experimental results on character recognition illustrate that the proposed method outperforms the other methods, involved in this study, by providing better recognition rates.

The cultural heritage is full of important manuscript collections preserved in digital libraries. The need to annotate and enrich the scanned documents is claimed by some users to keep traces in the system for a further use. Moreover, the reuse of annotations could help other users to accomplish repetitive tasks in a semi-automatic way. One manuscript annotation technique is the word spotting. It is a process that seeks in a document for all the fragments that are similar to the one specified by the user. The main focus of this research work is to propose a solution integrating and encapsulating the word spotting algorithm in digital libraries. This solution involves, in particular, the specification and the implementation of an architecture to integrate the image processing tool using Restful Web services. The proposed prototype is tested on the ARMARIUS digital library. This library is one of the collaborative digital archiving models that stores ancient digitized manuscripts.

Textual document image denoising is the main issue of this work. Therefore, we introduce a comparative study between two state-of-the-art denoising frameworks : local and non-local smoothing filters. The choice of both of these frameworks is directly related to their ability to deal with local data corruption and to process oriented patterns, a major characteristic of textual documents. Local smoothing filters incorporate anisotropic diffusion approaches where as non-local filters introduce non-local means. Experiments conducted on synthetic and real degraded document images illustrate the behaviour of the studied frameworks on the visual quality and even on the optical recognition accuracy rates.

A modification of the Weickert coherence enhancing diffusion filter is proposed for which new constraints formulated form the Perona-Malik equation are added. The new diffusion filter, driven by local tensors fields, takes benefit from both of these approaches and avoids problems known to affect them. This filter reinforces character discontinuity and eliminates the inherent problem of corner rounding while smoothing. Experiments conducted on degraded document images illustrate the effectiveness of the proposed method compared to another anisotropic diffusion approaches. A visual quality improvement is thus achieved on these images. Such improvement leads to a noticeable improvement of the OCR system's accuracy proven through the comparison of OCR recognition rates before and after the diffusion process.

Mean shift, a powerful color clustering approach successfully applied to image segmentation, has two main properties that are relevant for use in document image segmentation. These properties include: the autonomous definition of both color clusters' centers and numbers and the good tolerance to noisy data sets. Hence, mean shift could robustly process degraded background document images and improve their legibility. Nevertheless, this paper proves that coupling this approach and anisotropic diffusion within a joint iterative framework has more interesting results. For instance, this framework generates segmented images with more reduced artefacts on edges and background than those obtained after applying each method alone. This improvement is explained by the mutual interaction of global and local information, respectively introduced by the mean shift and anisotropic diffusion, and by the nature of this latter, smoothing while preserving continuities across edges. Some experiments, done on real ancient document images, illustrate these ideas and indicate that our proposed framework provides an efficient tool for document image segmentation and restoration.

This paper focuses on improving the optical character recognition (OCR) system 's accuracy by restoring damaged character through a PDE (Partial Differential Equation)-based approach. This approach, proposed by D. Tschumperle, is an anisotropic diffusion approach driven by local tensors fields. Actually, such approach has many useful properties that are relevant for use in character restoration. For instance, this approach is very appropriate for the processing of oriented patterns which are major characteristics of textual documents. It incorporates both edge enhancing diffusion that tends to preserve local structures during smoothing and coherence-enhancing diffusion that processes oriented structures by smoothing along the flow direction. Furthermore, this tensor diffusion-based approach compared to the existing sate of the art requires neither segmentation nor training steps. Some experiments, done on degraded document images, illustrate the performance of this PDE-based approach in improving both of the visual quality and the OCR accuracy rates for degraded document images.