Methods for improving texture description by using statistical information extracted from the image gradient
- García-Olalla Olivera, Oscar
- Enrique Alegre Gutiérrez Doktorvater/Doktormutter
- Laura Fernández Robles Co-Doktorvater/Doktormutter
Universität der Verteidigung: Universidad de León
Fecha de defensa: 29 von September von 2017
- Javier González Jiménez Präsident/in
- Víctor González Castro Sekretär/in
- Modesto Castrillón Santana Vocal
Art: Dissertation
Zusammenfassung
This thesis proposes three new descriptors developed to the same end of takingadvantage of statistical information extracted from the image gradient to improve the description of the texture present in an image. First, we proposed a new method, Adaptive Local Binary Patterns with oriented Standard deviation (ALBPS), based on the lack of information about the orientation in the well-known Local Binary Patterns (LBP). This idea was introduced previously in Adaptive Local Binary Patterns (ALBP), that takes into account the oriented mean and standard deviation of the local absolute difference in order to make the matching more robust against local spatial structure changes. Our proposal included the standard deviation information of the image gradient along different orientations not in the matching method but in the descriptor algorithm and it is called ALBPS on that account. Given the good results obtained by using ALBPS in comparison with LBP and most of its variants, ALBP included, we developed another new descriptor based on ALBPS that we named LOSIB (Local Oriented Statistical Information Booster). LOSIB is designed to improve the performance of other texture descriptors that mainly rely on local characteristics.It enhances description by extracting gray level differences along several orientations. Specifically, the mean of the differences along particular orientations is considered. Two additional parameters, i.e. the radius of the neighbourhood and the number of neighbours, give LOSIB more reliability and robustness than ALBPS, adapting the fetures extraction to different kinds of images and problems. Finally, we proposed CLOSIB (Complete LOSIB). It is a generalized version of LOSIB that incorporates three parameters into its configuration, i.e. the statistical order of the descriptor, the radius of the neighbourhood and the number of neighbours. CLOSIB is presented through a comprehensive mathematical expression and includes three different variants to provide a faster alternative, H-CLOSIB, a multi- scale descriptor, M-CLOSIB, and a combination of both, HM-CLOSIB. The main common feature of the proposed methods is that they retrieve global statistical information of the texture of the image in contrast to the local description provided by their predecessor LBP. We also used our methods in combination with other state-of-the-art local descriptors and the early fusion of both resulted in a higher discrimative power of the final descriptor over many experimental fields tested. Particularly, the concatenation of CLOSIB with methods based on Local Bin- ary Patterns (LBP) works fine because the characteristics used in their feature extraction are complementary. ALBPS was evaluated for material recognition and also for vitality assessment of boar spermatozoas. Regarding the first application, ALBPS achived a 61.47% of hit rate on KTH Tips2-a dataset outperforming all the other LBP-based methods evaluated. In relation to the latter field, ALBPS combined with 13 Wavelet Concurrent Features (WCF13) global descriptor, that extracts Haralick features from the Wavelet Transform, yielded up an F-Score of 0.886. This is the best result achieved so far in this field up to our knowledge. Three diverse application fields were evaluated to test the efficiency of LOSIB, which are materials recognition, classification of the acrosome integrity of boar spermatozoa and tool wear estimation. For materials recognition, two different datasets were evaluated: KTH-Tips-2a and Brodatz32 to prove the robustness of LOSIB. LOSIB was used together with LBP and several of its state-of-the-art variants. Combined with CLBP (Complete LBP), results were improved in 5.80% on KTH-Tips 2a and 7.09% on the Brodatz dataset. All tested descriptors in combination with LOSIB achieved a higher performance on both datasets using two different classifiers. Regarding acrosome integrity assessment of boar spermatozoa, three descriptors were combined: 13 Haralick features on the Gray Level Co-oncurrent Matrix (GLCM) of the Haar DWT; a texture description using LBP and a shape description based on Fourier. This method, called WFLP, was combined with LOSIB obtaining a 100% of accuracy and a 0.9975 of FScore, being able to precisely determine the state of the boar acrosomes without using neither stains nor expensive microscopes. For tool wear estimation, LOSIB combined wih ALBP achieved a 74.06% of hit rate for a binary classification of the wear level of milling inserts. CLOSIB was evaluated for three applications, i.e. material recognition –by means of KTH TIPS (2-a) dataset–, face recognition –using JAFFE dataset– and textile recognition –by means of creating and introducing a new dataset for textile retrieval in indoor rooms under challenging conditions of shape, illumination and occlusions–. The results showed that when we combine CLOSIB with different texture descriptors, the hit rate increases in all the cases, concluding that CLOSIB can be used to enhance the description of texture in a significant number of situations.