Dynamic Optimization Problems (DOPs) pose significant challenges because of the evolving nature of their objective functions and constraints over time. These difficulties become more pronounced as the frequency of landscape changes and the dimension of the search space increase. In this work, a novel hybrid approach for dynamic optimization, called Evolutionary Fractal Decomposition based Search (EFDS), is proposed. EFDS uses fractal-based decomposition for space indexing and Evolutionary Algorithms (EAs) to select prominent regions while maintaining population diversity. Experimental results on the Moving Peak Benchmark (MPB) demonstrate the effectiveness of the proposed approach, outperforming competing methods in 21 out of 24 benchmark configurations.

The ability of Transformers to process longer sequences has led to unprecedented levels of generalization in visual tasks. However, the complexity of Transformers is dominated by the quadratic cost associated with the computation of the attention blocks, posing a bottleneck that impedes the scaling of sequence length and the realization of more advanced AI solutions. We propose and explore the hypothesis that the self-attention mechanism exhibits regularities that can be exploited to enhance performance and achieve linear-cost attention without significant loss of effectiveness. Specifically, we investigate the attention matrix of Visual Transformers to identify and leverage these regularities in order to simplify the computation process. The resulting procedure significantly reduces the computational cost of Transformers by directly reducing attention block complexity. Moreover, the designed procedure is 0-shot self-supervised, thus it requires no retraining, additional data or parameters, as all Transformer parameters remain unchanged. Consequently, the proposed method can be seam-lessly applied to pre-trained Visual Transformers without the need for retraining. Experiments conducted on a series of Vision Transformers pre-trainedon ImageNet-1K dataset demonstrate the effectiveness of our proposed approach.

Dynamic optimization problems (DOPs) are known to be challenging due to the variability of their objective functions and constraints over time. The complexity of these problems increases further when the frequency of landscape change and the dimensionality of the search space are large. In this work, we propose a novel fractal decomposition-based method designed for DOPs, called FDS. It is a new single solution metaheuristic that introduces a new hypersphere-based space decomposition for efficient exploration, an archive for diversity control, and a pseudo-gradient-based local search (called GraILS) for fast exploitation. Extensive experiments on the well-known and the standard benchmark (the Moving Peak Benchmark: MPB) demonstrate that FDS consistently outperforms state-of-the-art competitors. Furthermore, FDS shows high robustness across diverse scenarios, maintaining superior performance despite variations in key benchmark parameters, such as the severity of landscape shifts, the number of peaks, the dimensionality of the problem, and the frequency of change. FDS achieves the highest average rank across all experiments and demonstrates dominant performance in 19 out of 23 scenarios. The implementation of FDS is available via the following GitHub repository: https://github.com/alc1218/FDS .

Abstract High-resolution computed tomography (CT) scans require high doses of X-rays, posing potential health risks to patients, including genetic damage and cancer. Conversely, low doses of X-rays result in noise and artifacts in the reconstructed CT scans. Consequently, the problem of denoising low-dose CT (LDCT) images has become a critical yet challenging issue in the field of CT imaging. However, existing deep learning-based LDCT image denoising methods frequently result in the loss of high-frequency features, such as edges and textures, due to the use of mean squared error loss. To address this issue, we propose a method based on high-frequency feature learning to enhance the denoising performance of existing models. Our method is designed to simultaneously learn the primary task of LDCT image denoising and the auxiliary task of LDCT edge detection, thereby improving the denoising performance without increasing the number of model parameters and the inference time. Our method significantly improves the denoising performance of the RED-CNN model, achieving competitive results compared to state-of-the-art denoising models on the AAPM and Qin-LUNG-CT datasets.

In the last decade, Convolutional Neural Networks became an industry standard achieving state-of-the-art results for many computer vision tasks. This unprecedented success has been possible due to the use of massive amounts of visual and multimodal data. However, management of these data must comply with the regulations on privacy protection, i.e. personal data should be anonymized. Traditional image anonymization methods such as blurring, masking, pixelating are efficient in the obfuscation of the sensitive data. Still, recent research has indicated that these methods impact in the negative way the performance of computer vision models. Numerous deep learning anonymization methods have been proposed as an alternative, in particular for human face and body anonymization. Unfortunately, the vast majority of these approaches are task-specific and require training. Other methods, although general, rely on full access to the computer vision model (the so-called white-box methods). Here, we propose a novel adaptive image anonymization method that allows one to achieve high concordance of the classification model predictions on the original and anonymized image. It is gradient-free, agnostic to anonymized objects, and to the particular architecture and weights of the computer vision model used. Finally, the proposed method does not require modifications to the computer vision model. The main idea of the approach introduced in this paper is to consider image anonymization as an optimization problem and to solve it using the iFDA metaheuristics algorithm. Experiments conducted on the large-scale benchmark image dataset ImageNet convincingly demonstrate the efficiency of our approach. When applying the proposed adaptive image anonymization method, the class concordance rate obtained was 98.11%, as opposed to 74.47% obtained by traditional anonymization.

Despite the evident practical importance of license plate recognition (LPR), corresponding research is limited by the volume of publicly available datasets due to privacy regulations such as the General Data Protection Regulation (GDPR). To address this challenge, synthetic data generation has emerged as a promising approach. In this paper, we propose to synthesize realistic license plates (LPs) using diffusion models, inspired by recent advances in image and video generation. In our experiments a diffusion model was successfully trained on a Ukrainian LP dataset, and 1000 synthetic images were generated for detailed analysis. Through manual classification and annotation of the generated images, we performed a thorough study of the model output, such as success rate, character distributions, and type of failures. Our contributions include experimental validation of the efficacy of diffusion models for LP synthesis, along with insights into the characteristics of the generated data. Furthermore, we have prepared a synthetic dataset consisting of 10,000 LP images, publicly available at https://zenodo.org/doi/10.5281/zenodo.13342102. Conducted experiments empirically confirm the usefulness of synthetic data for the LPR task. Despite the initial performance gap between the model trained with real and synthetic data, the expansion of the training data set with pseudolabeled synthetic data leads to an improvement in LPR accuracy by 3% compared to baseline.

Neural Networks (NNs) have shown excellent results in a variety of Machine Learning (ML) tasks and are now being used in a wide range of applications. Meanwhile, the size of these models has increased, with some of the most recent state-of-the-art models comprising billions of parameters. The necessity for compact and efficient NN representations has been recognized by research, which has provided specialized compression techniques for various applications. In this paper, we propose the optimization of the compression based on evolutionary algorithms and fractional differentiation. To this end, three main criteria were taken into account: filters error approximation using fractional differentiation, NNs accuracy and the Compression ratio. The results obtained on LeNet5 model demonstrated that there is no loss in terms of accuracy with 50% compression ratio in MNIST, 18% in CIFAR10, and 9% in CIFAR100. Moreover, the analysis showed that kernels of the first layer are more complex to be compressed than those of the second layer.

Deep learning based methods have become the de-facto standard for various computer vision tasks. Nevertheless, they have repeatedly shown their vulnerability to various form of input perturbations such as pixels modification, region anonymization, etc. which are closely related to the adversarial attacks. This research particularly addresses the case of image anonymization, which is significantly important to preserve privacy and hence to secure digitized form of personal information from being exposed and potentially misused by different services that have captured it for various purposes. However, applying anonymization causes the classifier to provide different class decisions before and after applying it and therefore reduces the classifier’s reliability and usability. In order to achieve a robust solution to this problem we propose a novel anonymization procedure that allows the existing classifiers to become class decision invariant on the anonymized images without any modification requires to apply on the classification models. We conduct numerous experiments on the popular ImageNet benchmark as well as on a large scale industrial toll classification problem’s dataset. Obtained results confirm the efficiency and effectiveness of the proposed method as it obtained 0% rate of class decision change for both datasets compared to 15.95% on ImageNet and 0.18% on toll dataset obtained by applying the naïve anonymization approaches. Moreover, it has shown a great potential to be applied to similar problems from different domains.

This paper proposes a hybrid strategy for accelerating image classification inference based on the Modern Continuous Hopfield Neural Network (MHNN). To implement this strategy, the fully connected layers of convolutional neural networks (CNNs) are replaced by the MHNN. The proposed hybrid architecture achieves promising results for image classification tasks, as demonstrated through experiments on multiple benchmark datasets, including ImageNet, and different CNN architectures. It offers a remarkable speedup in inference time (ranging from 1.12x to 1.6x) and significant compression in terms of the number of neural network parameters (ranging from 1.32x to 49.37x), while maintaining high accuracy. Furthermore, the proposed CNN-MHNN model achieves an accuracy of 99.18% on the Noisy MNIST dataset, outperforming state-of-the-art models with a 0.75% improvement for the Added White Gaussian Noise version.

The article focuses on the difficulties of training deep learning models for real-world scenarios, where the data is often noisy and improperly labeled. Despite the progress made in developing sophisticated models, they require extensive, well-labeled data sets to perform optimally. To tackle this challenge, various approaches have been proposed for training with noisy data, including our previous work, Co-Meta, which proposed a meta-learning approach. In this study, we employed Co-Meta to train a damage classification model with a noisy dataset. We demonstrate how the selection of meta data sets can significantly influence the results and present an efficient method for addressing this issue. Furthermore, we tackle the problem of high memory usage associated with Co-Meta by utilizing alternative bi-level optimization algorithms that decrease the computational resources needed. Our experiments demonstrate that the choice of optimization algorithm does not have a significant impact on the final accuracy, yet it has a considerable effect on the memory requirements. These modifications enhance Co-Meta performances significantly. Our selection strategy resulted in an accuracy improvement of over 1%, while the new optimization methods reduced the memory requirements by half.

Deep learning methods have shown state-of-the-art results in various application areas such as computer vision, NLP, etc. However, their practical use presents many challenges, including those caused by the large size of the models, especially in the context of model weight storage and transmission. One of the possible solutions to this problem is Neural Network (NN) compression, which is a process of obtaining a derived model serving the same task with a smaller number of parameters or with parameters of lower precision. The most common NN compression techniques include pruning, sparse representation, quantization, and knowledge transfer. In this article, the compression of Convolutional Neural Networks (CNNs) using fractional differentiation is investigated. A for-mulation of this task as a large-scale continuous optimization problem is then proposed, and its resolution is performed through a new optimization algorithm, called the Improved Fractal Decomposition Algorithm (IFDA), based on space geometric fractal decomposition. The results obtained show that MobileNetV3, for instance, is compressed by 18.5% with only a 2.5% decrease in accuracy. Additionally, the proposed IFDA algorithm outperforms all other competing metaheuristics in solving this problem.

Laws on privacy preservation challenges supervised learning algorithms in industrial applications and could be an obstacle for the artificial intelligence solutions. In the literature, this issue is never discussed for the algorithm’s design. Indeed, algorithms do not behave the same when the input is modified to protect privacy. Particularly, the unmodified data samples predicts with low confidences show high vulnerability to decision changes. To overcome this challenge, we propose a novel solution that enhances classifier’s robustness by particularly addressing the vulnerable samples. It consists of a novel formulation of the learning objective by hybridizing similarity learning, decision margin and intra-class distance. Experimental results and evaluation on a challenging vehicle image dataset exhibit the high effectiveness and potentials of our method for the privacy preserving classification problems.

This paper introduces the application of the fractal decomposition-based algorithm (FDA) to the optimization of the hyperparameters of deep neural network architecture. FDA is a metaheuristic that was recently proposed to solve high dimensional continuous optimization problems. In this work, we apply FDA to the optimization of well-known architectures such as VGG-16, NasNet, MobileNetV2 and ResNetV2-50. The hyperparameters of those architectures were fine-tuned using FDA on the CIFAR-10 benchmark dataset. The experiments demonstrate the superiority of proposed method over the state-of-art values. Considered approach shows promising results as every architecture was improved with hyperparameters found by using FDA. The experiments were conducted using low computational power with only 3 NVIDIA V100 GPUs, with 16GB of RAM.

In this paper, a real-world problem of the vehicle-type classification for automatic toll collection (ATC) is considered. This problem is very challenging because any loss of accuracy even of the order of 1% quickly turns into a significant economic loss. To deal with such a problem, many companies currently use optical sensors (OSs) and human observers to correct the classification errors. Herein, a novel vehicle classification method is proposed. It consists in regularizing the problem using one camera to obtain vehicle class probabilities using a set of convolutional neural networks (CNNs) and, then, uses the Gradient boosting-based classifier to fuse the continuous class probabilities with the discrete class labels obtained from the OS. The method is evaluated on a real-world dataset collected from the toll collection points of the VINCI Autoroutes French network. The results show that it performs significantly better than the existing ATC system and, hence, will vastly reduce the workload of human operators.

Image Quality (IQ) plays significant role for both human vision and artificial vision applications. This last decade, the number of camera increases exponentially, but the exploitation of the information depends on the quality of these acquired images, and sequences. A large number of objective blind IQ assessment (OBIQA) methods were proposed which find a global IQ estimation regardless the application. In this work, we assume that is impossible that such a method can be applied to all applications, however, for a given problem an objective model of the quality can be provided. Consequently, this paper addresses this issue for an automatic vehicle type classification application and proposes a novel OBIQA based classification approach. The proposed method first extracts a set of selective image features, then learns to classify images accordingly. In other terms, it aims to prevent misclassification and localize the source of poor images. Experiments show that it performs better than the state-of-the-art methods and can be used for similar applications.

Toll Vehicle Classification is an important task. Indeed, it has many uses in traffic management and toll collection systems. In this paper, Vinci Autoroutes group Networks (the biggest French Highways concession) are considered, where every year, millions of vehicles are classified in realtime. Then, a small decrease in classification performance can have serious economic losses. Therefore, the accuracy and the time complexity become critical for the toll collection system. The current classification algorithm uses the scene features' to detect vehicles classes. However, it requires a large labeled datasets, and has a limitations when multiple vehicles are in the scene. Herein, we propose a novel context-aware vehicle classification method that takes profit from the semantic spatial relationship of the objects. The experiments show that our method is performing as accurately as the existing model with significantly lower labeled datasets (74 times smaller). Moreover, the obtained accuracy of the proposed method is 99.97% compared to 99.79% achieved by the current method when using the same training set.

In this work, a challenging vehicle type classification problem for automatic toll collection task is considered, which is currently accomplished with an Optical Sensors (OS) and corrected manually. Indeed, the human operators are engaged to manually correct the OS misclassified vehicles by observing the images obtained from the camera. In this paper, we propose a novel vehicle classification algorithm, which first uses the camera images to obtain the vehicle class probabilities using several Convolutional Neural Networks (CNNs) models and then uses the Gradient Boosting based classifier to fuse the continuous class probabilities with the discrete class labels obtained from two optical sensors. We train and evaluate our method using a challenging dataset collected from the cameras of the toll collection points. Results show that our method performs significantly (98.22% compared to 75.11%) better than the existing automatic toll collection system and, hence will vastly reduce the workload of the human operators. Moreover, we provide an in-depth analysis w.r.t. the learning strategies:e.g., choice of the optimization algorithm of the CNN model. Our results and analysis highlights interesting perspectives and challenges for the future work.

International audience

Most range-free techniques for indoor localization depend on the received signal strength (RSS) fingerprints. Their performances are relied to the structure of the considered indoor environments. We consider in this paper RSS-based methods: Multi-Layer Perceptron Neural Network (MLPNN), and K-nearest neighbor (KNN), and compare their performance under the same indoor environment. One of the advantages focused by the choice of these techniques is their robustness against external disturbances that may affect the received RSS signal. Moreover, we propose a new metric to enhance the performance of the KNN method, called d-nearest neighbor. In order to test the different techniques, we build a heterogeneous fingerprint database with different resolutions. The obtained results show the efficiency of the proposed enhancement in the case of a heterogeneous high resolution database.

In this work, we deal with a class of problem of trajectory planning taking into account the smoothness of the trajectory. We assume that we have a set of positions in which the robot must pass. These positions are not assigned in the time axis. This kind of result can be found in many works of trajectory planning. Thus, this result is not complete in the sense that we do not have the complete trajectory and, we do not have any idea about the whole time to move. In this work, we propose a formulation of this problem, where the total length of the trajectory and the total time to move from the initial to the final position are minimized simultaneously. In order to avoid abrupt movement, we should ensure the smoothness of the trajectory at the position, the velocity and the acceleration levels. Thus, the position function must be at least two times differentiable. We use a genetic algorithm to resolve this problem and we show the efficiency of the proposed technique by simulation.

While the static indoor geo-location of mobile terminals (MTs) has been extensively studied in the last decade, the prediction of the trajectory of an MT is still a major problem when designing mobile location (tracking) systems (TSs). In fact, Global Positioning System (GPS) works quite well in outdoor conditions and relatively unobstructed spaces, but falls short in many urban conditions and other realistic use cases. It is important to augment mobile geo-location architectures with a prediction dimension to deal with distortions caused by obstacles, and ultimately produce a more accurate positioning system. Different prediction approaches have been proposed in the literature, the most common is based on prediction filters such as linear predictors (LPs), Kalman filters (KFs), and particle filters (PFs). In this paper, we take the prediction one step further by using digital fractional integration (DFI) to predict the actual trajectory of MTs. We evaluate the performance of our proposed DFI prediction in two indoor trajectory scenarios inspired by typical user mobility patterns in typical indoor conditions (museum visit and hospital doctor walk). To illustrate the efficiency of the proposed method in particularly noisy environments, we consider two other MT trajectory scenarios, namely spiral and sinusoidal trajectories. Experimental results show a significant performance improvement over most common predictors in the relevant literature, particularly in noisy cases. Extensive study of short-archive principle using 5, 10, and 25 previous estimated positions, showed the benefit of using DFI operator with only the most recent locations of an MT.

With the rise of Video-on-Demand (VoD) systems as a preferred way to distribute video content over IP networks, many research works and innovations have focused on improving the scalability of streaming systems by looking at distributed approaches such as peer-to-peer (P2P). One of the most critical aspects in P2P-assisted streaming system is the real-time resource allocation, which drives the performance of the system in terms of capacity utilization and VoD requests rejection rates. In this paper, we specifically focus on the problem of maximizing the P2P streaming system utilization by effectively alternating between different resource allocation strategies. Switching between different resource allocation strategies is guided by a run-time statistical analysis of performances against predicted content popularity pattern. A key contribution of this paper resides in effectively combining different, and potentially conflicting, performance objectives when deciding on which resource allocation strategy to use. Indeed, we use a Bayesian Fusion to select the most appropriate resource allocation strategy to deal with future content demand. With our P2P resource allocation framework, a VoD service operator can combine any number of resource allocation strategies and formulate different performance objectives that meet the requirements of its network and the content consumption behavior of its users.

Thin structure filtering is an important preprocessing task for the analysis of 2D and 3D bio-medical images in various contexts. We propose a filtering framework that relies on three approaches that are distinct and infrequently used together: linear, non-linear and non-local. This strategy, based on recent progress both in algorithmic/computational and methodological points of view, provides results that benefit from the advantages of each approach, while reducing their respective weaknesses. Its relevance is demonstrated by validations on 2D and 3D images.

During the past decade, Peer-to-Peer Video-on Demand (VoD) systems have proved their efficiency for large deployments. They raise new challenges such as peers resource allocation. Most literature on resource allocation tackle the problem with optimal static rules found at offline study of the system. In this paper, we use a dynamic metaheuristic, called Multiple Local-Search Algorithm for Dynamic Optimization (MLSDO) to optimize the problem at hand. The obtained results show that using a dynamic resource allocation reduces the rejection rate while enhancing the entropy of the system, in the face of a dynamically changing title demand.

Many real-world optimization problems are dynamic (time dependent) and require an algorithm that is able to continuously track a changing optimum over time. In this paper, we investigate a recently proposed algorithm for dynamic continuous optimization, called MLSDO (Multiple Local Search algorithm for Dynamic Optimization). MLSDO is based on several coordinated local search agents and on the archiving of the optima found over time. This archive is used when a change occurs in the objective function. The performance of the algorithm is evaluated on the set of benchmark functions provided for the IEEE WCCI-2012 Competition on Evolutionary Computation for Dynamic Optimization Problems.

Strong with large pools of set-top-boxes (STB) already deployed and connected to high-capacity networks, broadband operators sit on a major opportunity to introduce P2P-based VOD streaming architectures that scales up to the demand, reduce the per-service cost, and ultimately position the operators in a more competitive footing in respect to the increasingly successful Internet streaming services Achieving the required cost-efficiency in such services involves many research challenges, chief among which is the resource allocation in terms of assigning contributing STBs to stream a given VOD service. This paper1investigates resource allocation issues associated with the operation of P2P-based VOD streaming systems in the context of a broadband operator. First, we introduce basic resource allocation models that use a single-criterion (Available Uplink, Popularity Score, and Critical Score) to select contributing STBs and satisfy an incoming VOD request. We individually evaluate the performance of every single-criterion resource allocation strategy, and highlight its strengths and weaknesses in dealing with different situations.

While the static indoor location of a mobile terminal (MT) has been extensively studied on last decade, the prediction of the trajectory of a MT still is the major problem for building mobile location (tracking) systems (TSs). This problem is solved for outdoor TSs using global positioning system (GPS), however, it remains an essential obstacle to construct reliable indoor TSs. Different approaches were proposed in the literature, the most used is that based on prediction filters, such as linear filters (LF), Kalman filters (KF) and particle filters (PF). In this paper, we propose to enhance the performance of the predictors using digital fractional differentiation (DFD) to predict a MT trajectory. To illustrate the obtained results, three indoor trajectory scenarios inspired from real daily promenades are simulated (museum visit, hospital doctor walking and shopping in the market). Experimental results show a significant improvement of the performance of the classical predictors, particularly in noisy cases.

In this paper, we propose a multiagent based evolution strategy algorithm, called CMADO, to evaluate the amplitudes of the deformations of the walls of the third cerebral ventricle on a brain cine-MR imaging. CMADO based segmentation technique is applied on a 2D+t dataset to detect the contours of the region of interest (i.e. lamina terminalis). Then, the successive segmented contours are matched using a procedure of global alignment. Finally, local measurements of deformations are derived from the previously determined matched contours. The validation step is realized by comparing our results to the measurements achieved on the same patients through a manual segmentation provided by an expert using Ethovision ®software.

Many real-world problems are dynamic and require an optimization algorithm that is able to continuously track a changing optimum over time. In this paper, a new multiagent algorithm for solving dynamic problems is studied. This algorithm, called MADO, is analyzed using the Moving Peaks Benchmark, and its performances are compared to those of competing dynamic optimization algorithms on several instances of this benchmark. The obtained results show the efficiency of MADO, even in multimodal environments.