Jianfeng Song

RBoost: Label Noise-Robust Boosting Algorithm Based on a Nonconvex Loss Function and the Numerically Stable Base Learners

AdaBoost has attracted much attention in the machine learning community because of its excellent performance in combining weak classifiers into strong classifiers. However, AdaBoost tends to overfit to the noisy data in many applications. Accordingly, improving the antinoise ability of AdaBoost plays an important role in many applications. The sensitiveness to the noisy data of AdaBoost stems from the exponential loss function, which puts unrestricted penalties to the misclassified samples with very large margins. In this paper, we propose two boosting algorithms, referred to as RBoost1 and RBoost2, which are more robust to the noisy data compared with AdaBoost. RBoost1 and RBoost2 optimize a nonconvex loss function of the classification margin. Because the penalties to the misclassified samples are restricted to an amount less than one, RBoost1 and RBoost2 do not overfocus on the samples that are always misclassified by the previous base learners. Besides the loss function, at each boosting iteration, RBoost1 and RBoost2 use numerically stable ways to compute the base learners. These two improvements contribute to the robustness of the proposed algorithms to the noisy training and testing samples. Experimental results on the synthetic Gaussian data set, the UCI data sets, and a real malware behavior data set illustrate that the proposed RBoost1 and RBoost2 algorithms perform better when the training data sets contain noisy data.

Single image haze removal based on haze physical characteristics and adaptive sky region detection

Outdoor images are often degraded by haze and other inclement weather conditions, which affect both consumer photographs and computer vision applications severely. Therefore, researchers have proposed plenty of restoration approaches to deal with this problem. However, it is hard to tackle the color distortion problem in restored images with ignoring the differences between fog and haze. Meanwhile, the atmospheric light is also an important variable that influences the global illumination of images. In this paper, we analyze the physical meaning of atmospheric light first, and estimate atmospheric light by a novel method of obtaining the sky region in images, which is based on our newly proposed sky region prior. Then after exploring physical characteristics of fog and haze, we explain why images taken in haze appear yellowish, and eliminate this phenomenon by our adaptive channel equalization method. Quantitative comparisons with seven state-of-art algorithms on a variety of real-world haze images demonstrate that our algorithm can remove haze effectively and keep color fidelity better. HighlightsWe propose a novel single image haze removal approach.Our approach is based on haze physical character and adaptive sky region detection.We analyze the haze physical character and meaning of atmospheric scattering model.We propose a sky region prior based on thousands of outdoor images.We propose a new adaptive sky region detection method to estimate atmospheric light.We eliminate distortion in hazy images by our adaptive channel equalization method.We convert to HSI color space to compensate over-saturation phenomenon in RGB space.

Large-scale gesture recognition with a fusion of RGB-D data based on the C3D model

The gesture recognition has raised attention in computer vision owing to its many applications. However, video-based large-scale gesture recognition still faces many challenges, since many factors like background may disturb the accuracy. To achieve gesture recognition with large-scale videos, we propose a method based on RGB-D data. To learn gesture details better, the inputs are expanded into 32-frame videos first, and then the RGB and depth videos are sent to the C3D model to extract spatiotemporal features respectively. Next these features are combined to boost the performance, which can also avoid unreasonable synthetic data due to the uniform dimension of C3D features. Our approach achieves 49.2% accuracy on the validation subset of the Chalearn LAP IsoGD Database just with a linear SVM classifier. It also outperforms the baseline and other methods in the challenge and wins the first place at 56.9% on testing set.

Modular ensembles for one-class classification based on density analysis

One-Class Classification (OCC) is an important machine learning task. It studies a special classification problem that training samples from only one class, named target class, are available or reliable. Recently, various OCC algorithms have been proposed, however many of them do not adequately deal with multi-modality, multi-density, the noise and arbitrarily shaped distributions of the target class. In this paper, we propose a novel Density Based Modular Ensemble One-class Classifier (DBM-EOC) algorithm which is motivated by density analysis, divide-and-conquer method and ensemble learning. DBM-EOC first performs density analysis on training samples to obtain a minimal spanning tree using density characteristics of the target class. On this basis, DBM-EOC automatically identifies clusters, multi-density distributions and the noise in training samples using extreme value analysis. Then target samples are categorized into several groups called Local Dense Subset (LDS). Samples in each LDS are close to each other and their local densities are similar. A simple base OCC model e.g. the Gaussian estimator is built for each LDS afterwards. Finally all the base classifiers are modularly aggregated to construct the DBM-EOC model. We experimentally evaluate DBM-EOC with 6 state-of-art OCC algorithms on 5 synthetic datasets, 18 UCI benchmark datasets and the MNIST dataset. The results show that DBM-EOC outperforms other competitors in majority cases especially when the datasets are multi-modality, multi-density or noisy. We propose a modular ensemble OCC algorithm DBM-EOC based on density analysis.We analyze peculiarities of the target class which are crucial for OCC.DBM-EOC obtains a tree structure of the target class considering density.DBM-EOC can automatically detect clusters and remove noise samples.DBM-EOC solves OCC problems with the divide-and-conquer method.

Fast structural ensemble for One-Class Classification

We propose a fast structural ensemble OCC framework FS-EOCC.FS-EOCC does not need the determining of the number of clusters.FS-EOCC could convert a common OCC algorithm to a structural one.FS-EOCC run faster than most existing clustering based ensemble OCC methods. One of the most important issues of One-Class Classification (OCC) algorithm is how to capture the characteristics of the positive class. Existing structural or clustering based ensemble OCC algorithms build description models for every cluster of the training dataset. However, the introduction of clustering algorithm also causes some problems, such as the determination of the number of clusters and the additional computational complexity. In this paper, we propose Fast Structural Ensemble One-Class Classifier (FS-EOCC) which is a fast framework for converting a common OCC algorithm to structural ensemble OCC algorithm. FS-EOCC adopts two rounds of complementary clustering with fixed number of clusters. This number is calculated according to the number of training samples and the complexity of the base OCC algorithm. Each partition found in the previous step is used to train one base OCC model. Finally all base models are modularly aggregated to build the structural OCC model. Experimental results show that FS-EOCC outperforms existing structural or clustering based OCC algorithms and state-of-the-art non-structural OCC algorithms. The comparison of running time for these algorithms indicates that FS-EOCC is an efficient framework because the cost of converting a common OCC algorithm to a structural OCC algorithm is small and acceptable.

The Recognition of the Point Symbols in the Scanned Topographic Maps

It is difficult to separate the point symbols from the scanned topographic maps accurately, which brings challenges for the recognition of the point symbols. In this paper, based on the framework of generalized Hough transform (GHT), we propose a new algorithm, which is named shear line segment GHT (SLS-GHT), to recognize the point symbols directly in the scanned topographic maps. SLS-GHT combines the line segment GHT (LS-GHT) and the shear transformation. On the one hand, LS-GHT is proposed to represent the features of the point symbols more completely. Its R-table has double level indices, the first one is the color information of the point symbols, and the other is the slope of the line segment connected a pair of the skeleton points. On the other hand, the shear transformation is introduced to increase the directional features of the point symbols; it can make up for the directional limitation of LS-GHT indirectly. In this way, the point symbols are detected in a series of the sheared maps by LS-GHT, and the final optimal coordinates of the setpoints are gotten from a series of the recognition results. SLS-GHT detects the point symbols directly in the scanned topographic maps, totally different from the traditional pattern of extraction before recognition. Moreover, several experiments demonstrate that the proposed method allows improved recognition in complex scenes than the existing methods.

Color topographical map segmentation Algorithm based on linear element features

In order to overcome the discontinuity of geographic elements during the digitization of scanned topographic maps, a color map segmentation algorithm, which is used to segment color maps into different layers based on linear element features, is proposed in this paper. Linear elements are regarded as the elementary units in this method. We use background removal, thinning, nodes disconnection, labeling and dilation to get the elementary units. Then the main color, which could accurately represent the color feature of linear element, is extracted for clustering on the basis of Fuzzy c-means algorithm. At last, disconnected nodes are merged into the corresponding layers to keep the continuity of the results. The experimental results show that the proposed algorithm outperforms other segmentation approaches that regarding pixels as the elementary units.

SCTMS: Superpixel based color topographic map segmentation method☆

Abstract Different from natural image, topographic map is a complex manually generated image which has amount of interlaced lines and area features. Because of the frequent intersection and the overlap between geographic elements, the misalignment in scanner and other disturbances like inappropriate preserving, false color, mixed color and color aliasing problems occur in the raster color maps. These problems could cause serious challenges in segmentation process. In this work, we present a color topographic map segmentation method based on superpixel to overcome these problems. Firstly, the finest partition is obtained based on double color-opponent boundary detection method and watershed approach. Then, a strict region merging method is introduced to prevent mis-merging while superpixels generated. This merging method could make the superpixel partition accurately adherent the boundary between different geographic elements. Finally, luminosity, color and texture information are combinative applied to classify the superpixel into different layers based on support vector machine. The experimental results show that the proposed method outperforms other state-of-art topographic map segmentation approaches.

Improved road centerlines extraction in high-resolution remote sensing images using shear transform, directional morphological filtering and enhanced broken lines connection

Road information plays an important role in many civilian and military applications. This paper proposes an improved method for road centerlines extraction, which is based on shear transform, directional segmentation, shape features filtering, directional morphological filtering, tensor voting, multivariate adaptive regression splines (MARS) and enhanced broken lines connection. The proposed method consists of five steps. Firstly, directional segmentation based on spectral information and shear transform is used to segment the images for obtaining the initial road map. Shear transform is introduced to overcome the disadvantage of the loss of the road segment information. Secondly, we perform hole filling to remove the holes due to noise in some road regions. Thirdly, reliable road segments are extracted by road shape features and directional morphological filtering. Directional morphological filtering can separate road from the neighboring non-road objects to ensure the independence of each road target candidate. Fourthly, tensor voting and MARS are exploited to extract smooth road centerlines, which overcome the shortcoming that the road centerlines extracted by the thinning algorithm have many spurs. Finally, we propose an enhanced broken lines connection algorithm to generate a complete road network, in which a new measure function is constructed and spectral similarity is introduced.Display Omitted We apply spectral information and shear transform in directional segmentation.Directional morphological filtering is adopted to ensure the independence of road.We propose an enhanced broken lines connection algorithm. Road information plays an important role in many civilian and military applications. Road centerlines extraction from high-resolution remote sensing images can be used to update a transportation database. However, it is difficult to extract a complete road network from high-resolution images, especially when the color of road is close to that of background. This paper proposes an improved method for road centerlines extraction, which is based on shear transform, directional segmentation, shape features filtering, directional morphological filtering, tensor voting, multivariate adaptive regression splines (MARS) and enhanced broken lines connection. The proposed method consists of five steps. Firstly, directional segmentation based on spectral information and shear transform is used to segment the images for obtaining the initial road map. Shear transform is introduced to overcome the disadvantage of the loss of the road segment information. Secondly, we perform hole filling to remove the holes due to noise in some road regions. Thirdly, reliable road segments are extracted by road shape features and directional morphological filtering. Directional morphological filtering can separate road from the neighboring non-road objects to ensure the independence of each road target candidate. Fourthly, tensor voting and MARS are exploited to extract smooth road centerlines, which overcome the shortcoming that the road centerlines extracted by the thinning algorithm have many spurs. Finally, we propose an enhanced broken lines connection algorithm to generate a complete road network, in which a new measure function is constructed and spectral similarity is introduced. We evaluate the performance on the high-resolution aerial and QuickBird satellite images. The results demonstrate that the proposed method is promising.

Remote Sensing Image Fusion Based on Shearlet and Genetic Algorithm

Image fusion is a technology which can effectively enhance the utilization ratio of image information, the accuracy of target recognition and the interpretation ability of image. However, traditional fusion methods may lead to the information loss and image distortion. Hence a novel remote sensing image fusion method is proposed in this paper. As one of the multi-scale geometric analysis tools, Shearlet has been widely used in image processing. In this paper, Shearlet is used to decompose the image. Genetic Algorithm, a intelligent optimization algorithm, is also applied to image fusion and it aims to optimize the weighted factors in order to improve the quality of fusion. Experimental results prove the superiority and feasibility of this method.