Multi-view semi-supervised classification (Mv-SSC) aims to leverage complementary information from multiple views and unlabeled data to enhance classification performance. However, existing methods often struggle with constructing robust graphs, effectively fusing heterogeneous features. Therefore, we propose a novel framework, Multi-view Semi-supervised Classification with Graph Construction Innovation and Smoothness-aware Graph Convolution (GCSGC), which introduces three key innovations. First, we employ sparse autoencoders (SAEs) coupled with a learnable feature fusion module based on a one-layer fully connected network. The SAEs extract compact and discriminative representations from each view by enforcing sparsity to remove redundancies, while the fully connected layer adaptively aligns the multi-view latent features with the learnable shared features to preserve complementary semantic information. Second, GCSGC introduces two types of view-based graphs: (1) a Cosine-KNN-based Collaborative Graph (CKG) that integrates K-Nearest Neighbor relationships with cosine similarity of the learned view-specific latent features, and (2) a robust semi-supervised graph that enhances inter-view consistency and structural stability. Third, we design a hybrid loss function that combines cross-entropy and smoothness regularization to guide the optimization of the graph convolutional network. Extensive experiments on multiple benchmark datasets demonstrate that GCSGC consistently outperforms state-of-the-art Mv-SSC methods, validating the effectiveness of its innovative graph construction strategy, sparse feature fusion mechanism, and dual-loss optimization. This work provides a new perspective on integrating structural modeling, feature transformation, and loss design in multi-view semi-supervised learning.

Recently, null proximal discriminant analysis (NPDA) has been introduced in [1] to identify anomalies in a target data set. However, the proposed algorithm involves high computational burden and memory requirement which limit its use to moderate sized static datasets. The goal of this paper is to propose an incremental and sparse NPDA which is able both to process sequentially a large volume of data and remove the redundant information. More precisely, our solution lies in computing the null projection direction by a recursive singular value decomposition in the kernel induced feature space. Sparsity of the solution is achieved by thresholding the null score between the already processed data and newly injected data. The combined action of these two steps introduces a significant gain in terms of time complexity and memory burden while offering comparable classification performance as the batch method. Experimental results on large-scale data sets confirm the effectiveness of the proposed algorithm.

Global illumination methods are used to simulate lighting in 3D scenes. They provide a progressive convergence to high quality photo-realistic images as proved by Monte Carlo theory. One of the problem of such methods is to determine a stopping condition in order to decide if the computation reaches a satisfactory convergence which allows the process to terminate. In this paper, an inductive learning model for a reduced-reference quality assessment model in large-scale images is proposed as a solution for this problem. One key issue for learning algorithms in global illumination is that they are very efficient to learn perceptual features but are less efficient to learn stochastic noise. Moreover, they need a complete framework with a huge number of images in order to train and evaluate the learning model. These images are very difficult to obtain because of the time required for modeling and scene rendering. The idea is to improve performance by selecting the most pertinent images for the noise perception algorithm. In order to generalize the performance of our approach, experiments are conducted on global illumination scenes, with different precision and noise levels, computed with diffuse and specular rendering. Compared with human psychovisual scores, it can be seen the good consistency between these scores and the learning model quality measures. The proposed model has also been compared with the SVM learning model. The obtained results show that the proposed method is powerful to generalize learning and gives promising efficiency and precision.

Analysis of lace texture images is a challenging problem because the lace is a soft and extensible material and can be easily deformed. This paper investigates a whole system for lace classification. A first step, based on Otsu's segmentation method, allows to remove the background. Then the lace texture is characterized using local binary patterns (LBP). In order to be robust against rotation the Fourier Transform is applied on LBP histograms. The magnitude spectrum of this transform is then used as a feature vector. LBP descriptor parameters, including radius and number of neighbors, are adjusted in order to improve their relevance. The experiments show that the features based on LBP, with appropriate settings, produced good results in supervised and unsupervised contexts.

The images of lace textile are particularly difficult to be analyzed in digital form using classical image processing techniques. The major reasons of this difficulty emerge from the complex nature of lace which generally has different textures in its constituents like the background and patterns. In this paper, we study the behavior of Image Histogram (HistI) and Local Binary Patterns (LBP) on image extracts of lace in presence and absence of rotation. We further evaluate two variants of LBP; primarily the LBP Histogram (LBPB) and secondly the Fourier Transform applied on the LBP Histogram (LBPFFT). Consequently, we analyze the contribution of data fusion on feature level and score level in the different experimentations. The classification rate evaluates the discrimination degree of each descriptor via the k nearest neighbors kNN classifier. Experimental results indicate that the LBPB, LBPFFT and HistI combined at score level generate the better performance in absence of transformations. Whereas, LBPFFT and HistI combined at the same level generate the better classification rate, in the presence of rotation.

In this paper, we propose a Local Binary Pattern (LBP) histogram selection approach. It consists in assigning to each histogram a score which measures its efficiency to characterize the similarity of the textures within the different classes. The histograms are then ranked according to the proposed score and the most discriminant ones are selected. Experiments, which have been carried out on benchmark color texture image databases, show that the proposed histogram selection approach is able to improve the classification performances.

The paper deals with a monitoring system combining K-means classifier and one Hidden Markov Model in order to detect phytoplankton blooms and to understand their dynamics. The states of the Hidden Markov Model and codebook symbols are computed without a priori knowledge thanks to K-means algorithms. The system is tested on database signals from the Marel-Carnot station that registers water characteristics at high frequency resolution. The experiments show that, when the states are set to two, these correspond to phytoplankton productive and non-productive periods. Moreover, when states are set to five, these correspond to the dynamics of phytoplankton blooms.

Support vector regression (SVR) is now a well-established method for estimating real-valued functions. However, the standard SVR is not effective to deal with severe outlier contamination of both response and predictor variables commonly encountered in numerous real applications. In this paper, we present a bounded influence SVR, which downweights the influence of outliers in all the regression variables. The proposed approach adopts an adaptive weighting strategy, which is based on both a robust adaptive scale estimator for large regression residuals and the statistic of a "kernelized" hat matrix for leverage point removal. Thus, our algorithm has the ability to accurately extract the dominant subset in corrupted data sets. Simulated linear and nonlinear data sets show the robustness of our algorithm against outliers. Last, chemical and astronomical data sets that exhibit severe outlier contamination are used to demonstrate the performance of the proposed approach in real situations.

Support Vector Regression (SVR) is now a well-established method for estimating real-valued functions. However, the standard SVR is not effective to deal with outliers and structured outliers in training data sets commonly encountered in computer vision applications. In this paper, we present a weighted version of SVM for regression. The proposed approach introduces an adaptive binary function that allows a dominant model from a degraded training dataset to be extracted. This binary function progressively separates inliers from outliers following a one-against-all decomposition. Experimental tests show the high robustness of the proposed approach against outliers and residual structured outliers. Next, we validate our algorithm for object tracking and for optic flow estimation.

Support Vector Regression (SVR) is now a well-established method for estimating real-valued functions. However, the standard SVR is not effective to deal with outliers and structured outliers in training data sets commonly encoutered in computer vision applications. In this paper, we present a weighted version of SVM for regression. The proposed approach introduces an adaptive binary function that allows a dominant model from a degraded training dataset to be extracted. This binary function progressively separates inliers from outliers following a one-against-all decomposition. Experimental tests show the high robustness of the proposed approach against outliers and residual structured outliers. Next, we apply the algorithm to motion estimation in cluttering backgrounds with very encouraging results.

In this paper, we present a new support vector regression (SVR) based strategy for simultaneously extracting multiple linear structures in a training data set. As in fuzzy c-prototypes algorithms [17], [18], [10], we introduce fuzzy weights in the SVR formulation which assign to each data point a membership value according to c-structures. We propose to solve the corresponding dual problem under an iterative strategy with an initialization step. Experiments show the benefits of robustness properties of SVR in comparison with the standard fuzzy c-prototypes algorithm. Next, the motion estimation problem is used to illustrate its applicability and relevance in respect of real-world applications.

The robust regression is an important tool for the analysis of data contamined by outliers. In computer vision, the optic flow computation is considered as belonging to this kind of problem. In this paper, we discuss a robust optic flow computation based on a modified support vector regression (SVR) technique. We experimentally show that the proposed method significantly improves the robustness against outliers compared to traditional SVR. Next, we illustrate the performances of the method for the optic flow computation problem from noised image sequences. 1

Clustering has been known as a popular technique for pattern recognition, image processing, and data mining. Unfortunately, all known clustering algorithms tend to break down in high dimensional spaces; this is due to the inherent sparsity of the points. We investigate, in this paper, the use of linear and nonlinear principal manifolds for learning low-dimensional representations for clustering. Several leading methods: PCA, KPCA, Sammon, and CCA are examined and tested in clustering experiments using synthetic and real datasets from the UCI databases. We compare the clustering performance of the K-means algorithm on data projected by these projection methods. The experimental results show that K-means clustering on data projected by KPCA outperforms those projected by the three other methods

This paper describes a vision system that detects cracks in glass bottles production. The first step consists in collecting prototypes of bottles with and without defects. A sequence of 16 images is captured by a matrix camera while each bottle rotates in front of a specific lighting system. The second step is concerned with morphometric and photometric features extraction. The subsequent decision step is performed by different neural networks, such as MLP, RBF, PNN and LVQ. Finally, performances of these networks have been compared. All the images of bottles without defects have been recognized but a few images with small cracks, which are very important defects, have not been identified. However, since each bottle is represented by a sequence of 16 images, cracks will appear in at least three or four images, so that a defective bottle can be detected at least one time through the sequence. Therefore the decision system recognizes good and defective bottles with a very high rate of success.

In glass bottle inspection, the defects detection is of first importance. For online system detection, high speed and robust detection of faults are highly required. Neural networks have recently, and successfully, been applied to fault detection in many manufacturing processes. In this study, a Gaussian neural network, an extension of the RBF network, trained through a competitive algorithm, has been chosen for fault detection. Four parameters extracted from images of the bottles are used as inputs of the network. The number of Gaussian units is adjusted by an informational criterion. Experimental results show that the performance of this network are better than classical parametric and non parametric classifiers.