The Researches And Improvements Of Texture Features Extraction Algorithms Based On Digital Images

For now, pattern recognition is widely used in the biological security, automatic pro-duction, Internet payment and many other aspects, mode features are the observations of mode samples, the feature extraction algorithms have a crucial impact on the performance of the whole recognition system. There are obvious differences between different pattern categories and some similarities in the same category for a better features extraction al-gorithm. Researches are carried out on how to improve the features representativeness of objects or target images, some feature extraction algorithms are improved in this paper.For the color features extraction algorithms, according to the shortage of the orig-inal color histogram algorithm in some special cases, a weighted improved algorithm is proposed in this paper, the features which are extracted using improved algorithm have more representativeness and could get more representative color features in any situation.For the shape features extraction algorithms, combined with the characteristics of different degree of circularity, the aspect of object coverage rectangle is proposed that the description of objects shape features is more proper, some other object shape feature extraction algorithms are expounded and summarized in the paper.For the texture features extraction algorithms, the gray level co-occurrence matrix algorithm and texture spectrum algorithm are researched. For extraction error of pixel pairs in GLCM, the aspect of "Faked Pixels" is proposed, the accuracy of pixel pairs gray values is improved; four other directions are added and the rotation invariance of texture features is improved further. Against that the texture spectrum algorithm has no rotation invariance, more arrangement modes of pixels are increased; the gray value levels are increased and the improved texture unit has better texture description ability.Improved texture feature extraction algorithms are verified by using Brodatz im-ages dataset, combining with BP neural network and SVM, the recognition accuracy is compared and analyzed.The experimental results show that the two improved algorithms in this paper have higher recognition accuracy and better rotation invariance compared with the original GLCM algorithm; compared with the original texture spectrum, the first improved algo-rithm has some rotation invariance and the second one has better description ability and higher recognition accuracy.