Superpixels And Fuzzy Feature Learning Based Image Classification And Recognition

With the development of computer science and the popularity of the Internet,there are more and more images available in real life.Huge amounts of image data has brought more rich and intuitive information.However,large amount of images raises a series of problems.One of the important problem is how to distinguish between different images and pick up useful image from the massive image data.The problem has promoted the study on image classification and recognition.Nowadays,universities,research institutes and companies have studied the image classification problem.Image preprocessing and feature extraction are two key technologies affect the performance of image classification.Superpixels are often used as a preprocessing step for image processing technology.As superpixels consider both the spatial relationship between pixels and preserve the local characteristics of the image,superpixels-based classification methods have attracted much more attention than pixelsbased classification methods.Besides,the hundreds of superpixels instead of thousands of pixels in superpixels-based method,which improve the computation efficiency of the algorithm.Rough set theory can be used to the data analysis and reasoning directly,and also can be used as a data preprocessing method combined with other algorithms(such as neural networks).The paper mainly study the image classification and recognition problem from two aspects:(1)Superpixels,study superpixels for image classification problem;(2)feature learning,how to extract useful feature by rough set theory and neural network.The main contribution of this paper have been summarized as follows.1.Rough set theory has been proved to be an excellent mathematical tool for feature extraction.As the feature subset obtained by rough set theory is not unique,rough set theory can be integrated with ensemble learning.Ensemble learning has been a hot topic in machine learning due to its successful utilization in many applications.It is commonly agreed that the success of ensembles is attributed to two points.First,the accuracy of base classifiers is not particularly good but slightly better than random guess.Second,the diversity among the base classifiers.Therefore,a dynamic rough subspace based selective ensemble algorithm is proposed.The feature subsets obtained by rough set theory have the same discernible power with the original dataset.Theoretically,the performance of the base classifiers trained by the feature subsets is not bad.In order to improve the diversity among base classifiers,different feature subsets are necessary.In proposed algorithm,the relationship among attributes in rough subspace is first considered,and the maximum dependency degree of attribute is first employed to effectively reduce the searching space of feature subsets and augment the diversity of selected feature subsets.In addition,in order to choose an appropriate feature subset from the dynamic feature subsets searching space,an assessment function which can balance the accuracy and diversity is utilized.An ensemble system is built by base classifiers trained by these select feature subsets.Experimental results with UCI datasets and face images demonstrate that the proposed algorithm can lead to a comparative or even better performance in classification problem.2.Due to the limitation of rough set theory,the describing ability the features reduced by rough set theory is limited.However,rough set theory can be used as a preprocessing tool and combine with other soft computing tool.A novel two-layer feature learning framework is proposed to address single sample per person face recognition.The framework combine fuzzy rough set theory with sparse autoencoder.Fuzzy rough set is used to select features from original features,which enhance computer velocity.Sparse autoencoder is used to extract the inner structure of features,which improve the describing ability of rough set theory.In the first layer,the objective-oriented local features are learnt by a patch-based fuzzy rough set feature selection strategy.The obtained local features are not only robust to the image variations,but also usable to preserve the discrimination ability of original patches.Global structural information is extracted from local features by a sparse autoencoder in the second layer,which reduces the negative effect of non-discriminative regions.Besides,the proposed framework is a shallow network,which avoids the over-fitting caused by using multi-layer network to address single sample per person problem.The experimental results on ten face datasets have shown that the proposed local-to-global feature learning framework can achieve superior performance than other state-of-the-art feature learning algorithms for single sample per person face recognition.3.As superpixels-based image processing technology have many advantages than pixels-based image processing technology,superpixels methods have drawn much attention in recently years.Each superpixels algorithm has its own advantages.Different superpixels algorithms are suitable for different application problems.Nowadays,no superpixel algorithm is especially designed for image classification.In classification,all pixels in a superpixel are regarded as having the same labels.However,it is believed that both mixed superpixels and pure superpixels exist in practice applications.Pure superpixels are composed of pixels from a single class,and mixed superpixels consist of pixels with various labels.As mixed superpixels have negative effects on classification accuracy,a novel superpixels concept,named fuzzy-superpixels,is proposed.Pixels in an image are divided into two parts according to fuzzy-superpixels,superpixels and undetermined pixels.That is,not all pixels are assigned to a specific superpixel.Based on the concept of fuzzy-superpixels,we would rather ignore the pixels than assign them to an improper superpixel.Second,an algorithm,named FS,is presented to generate fuzzy-superpixels for Pol SAR images classification.Experimental results on three Pol SAR images demonstrate the superiority of the proposed fuzzy-superpixels algorithm over several state-of-the-art superpixels algorithms.4.Fuzzy equivalent relation based fuzzy-superpixels algorithm is proposed for Pol SAR image classification.Fuzz equivalent relation is a usually used concept in rough set theory,which can be used to measure the dependence between features.In this paper,the features in Pol SAR image are measured by fuzzy equivalent relation.The features of pixels in the same class have high dependence.Besides,according to fuzzy equivalent relation,the proportion of undetermined pixels can be adjusted adaptively.In the proposed algorithm,first,initialize the cluster centers and find out non-overlapping search region and overlapping search region.Second,pixels in overlapping search region belong to the superpixels of the corresponding cluster centers.For pixels in overlapping search region,whether a pixel belonging to a corresponding superpixel need to be judged.At last,execute the post-processing step to enforce the region connectivity.Experiments on three Pol SAR images show that the proposed algorithm reduces the number of mixed superpixels,and the cost of boundary adherence.