Research On Algorithm And Application Of Feature Extraction Based On Image Texture

Texture, one of basic attributes of the image and the descriptions of intrinsic property for object surface, provides important visual cues for image analysis. Texture classification is an important topic in the fields of image processing, computer vision and pattern recognition, and has received a lot of research interest and attention during the past decades. There are a wide variety of potential applications, such as remote sensing analysis, industrial fabrics inspection, medical image analysis, scene and object recognition, content-based image and video analysis, and material classification. However, the major challenge for texture classification, especially for analyzing real-world textures, lies in the large variety of geometric, stochastic and photometric transformations on the appearance of textures, which is caused by illumination changes, rotation variations, variability in scale and contrast, viewpoint changes, non-rigid deformations and occlusions. Therefore, this paper focus on how to extract effective texture features which are able to highly distinguish between intra-class and inter-class attributes of texture. It means that the constructed texture descriptors can maximize the similarity of inter-class while minimizing the ones of intra-class. The main contents of this paper are as follows.For the problem of sensitiveness to noise, high dimensionality and inadequate description of macro-mode existing in the Local binary pattern (LBP) descriptor, an effective sampling structure based on Pixel To Patch (PTP) to mimic the retinal sampling pattern and a novel local neighboring intensity relationship pattern (LNIRP) descriptor are proposed. The proposed PTP sampling structure simultaneously captures micro-patterns and macro-patterns. LNIRP descriptor, which is complementary to the LBP descriptor, is built by using neighboring intensity relationship (NIR) operator to explore neighboring gray-scale properties. A new texture description method is proposed based on PTP and NIR. With this method, LNIRP and LBP features are firstly fused jointly. Then, the joint descriptors are extended by using PTP sampling structure to describe textures. Theoretical analysis and texture classification experimental results show that the proposed descriptor has advantages of robustness to noise, low feature dimensionality, higher computational efficiency and classification accuracy. Two novel operators, circum-directional derivative (CDD) and circum-directional pattern (CDP)--are proposed for gray-scale and rotation invariant texture classification. And a new texture description method is proposed based on the above two operators. Firstly, it adopts CDD operator to explore circum-directional derivative information, and CDP operator to capture the central pixel information and spatial structure (pattern) information between central pixel and circum-directional neighboring pixels. Secondly, CDD and CDP operators are further extended to higher orders and different variants to encode more discriminative information in a given local region respectively. Thirdly, in a way similar to LBP, these two operators are then adopted to build two new texture descriptors respectively, namely CDD for the local binary circum-directional derivative descriptor (LB-CDD) and CDP for the local binary circum-directional pattern descriptor (LB-CDP). Finally, the proposed LB-CDD and LB-CDP descriptors are fused jointly. According to the different orders of LB-CDD and different variants of LB-CDP, several fused descriptors are built. Based on several challenging rotation invariant texture classification databases, experimental results show that the proposed method significantly outperformed other methods in classification accuracy as well as keeping a smaller feature dimension.For the problem of high expenses for computation and storage space of Gabor transformation based facial expression description, a combination of PTP based sampling structure and spatial saliency based method is proposed for the description of facial expression. It firstly adopts monogenic signal analysis to decompose a facial image into three feature maps for monogenic amplitude, phase and orientation. Secondly, each feature map is divided into multiple rectangular sub-regions. The amplitudes of the rectangular sub-region are set as the spatial saliency which is allocated to each rectangle as different weights. Then, on the each rectangular sub-region, LBP feature based on PTP are respectively extracted from three monogenic maps. Finally, the combination of those features, which are weighted by spatial saliency, are concatenated together to further enhance the discrimination. Experimental results on facial expression databases show that the proposed method which has higher accuracy and lower feature dimension, is an effective method of facial expression recognition.For the problem of low efficiency and accuracy for fastening systems inspection in safe rail transportation, a method on fastener detection and status inspection is proposed based on PTP sampling structure and visual tracking technology. Firstly, coarse positions of fasteners can be detected through adopting local region’s average intensity and projection residuals of gray-values. The positions of detected sleepers are set as coordinates of fasteners. Based on PTP, LBP features are extracted from positive and negative samples which are sampled around the fasteners. The exact locations of fasteners are tracked via naive Bayes classifier capable of online learning and automatic updating. Finally, the features extracted from exact location of fasteners are classified to detect the status of the fasteners. Adopting the railway video databases, it experimentally showed that the proposed approach can effectively track the fasteners and automatic inspect their status. The proposed method which has higher recognition accuracy and real-time tracking speed, is an effective and robust inspection approach.Finally, after analyzing the relevant theories and methods, we conclude the proposed multi-scale texture feature extraction method and its application in face expression recognition and railway fastener detection. For the shortcoming of the presented method, we discuss and analyze the further research work in the future.