Study On The Algorithm Of Image Feature Extraction And Its Application

Feature extraction is a significant step in object detection and recognition based on images. More than one feature could be found with an image such as appearance, statistic feature of pixels, coefficient feature with transformed image, and algebraic feature. These different features all can represent and distinguish one image from another. But they express the different character of one image. So the various feature could be used for different purposes with different objects. There are further researches to do on how to select an efficient method of feature extraction to satisfy the application demand, how to improve the performance of object detection and recognition with existing methods and so on.After analysis of the existing methods of feature extraction, more researches have been done on feature extraction for its application in face detection, face recognition and flame detection. Some more efficient algorithms of feature extraction are developed in this paper.The method of fast face detection based on AdaBoost and Cascade algorithms is popular because of its real time processing. But the training time consumption would be very expensive to get a face detector using AdaBoost algorithm. The expensive time consumption prevented AdaBoost being used widely. To solve the problem, the idea of feature-value-division is presented, and the improved AdaBoost algorithm based on feature-value-division is developed. Furthermore, the weight histogram of face and nonface are computed by using these quantified feature values. The threshold of simple classifier with each feature is computed faster based on the weight histogram. According to the distribution of feature values of face and nonface, dual-threshold is defined and computed. Then the improved algorithm called DW-AdaBoost is proposed based on Dual-threshold with the weight histogram. Experimental results on MIT-CBCL face and nonface training data set illustrate that the DW-AdaBoost algorithm could make training process convergence quick and effective. More experimental results on MIT+CMU using the detector also show that the detection speed and detection precision both reach higher performance.Linear discriminant analysis (LDA) is a classical and widely used technique for feature extraction. The feature extracted from the initial image based on LDA is easily influenced by the light condition, noise, and rotation of the face on the image. There are also some local characteristics lost after only the general feature is extracted. To avoid these problems, the new idea is put forward that the feature extraction would be done after the image is divided into multi-scale modular and transformed into other forms. Multi-scale Low Frequency Linear Discriminant Analysis (MSLF+LDA) and Multi-scale SVD Linear Discriminant Analysis (MSVD+LDA) are developed respectively. A series of experiments have been performed. The experimental results on ORL human face data set illustrate that the performance of the proposed method is obviously improved, moreover the dimension of the initial feature data is dropped down also obviously.Then the MSLF and MSVD are fused for face recognition named Multi-scale Multi-feature Fusion Linear Discriminant Analysis (MMFF+LDA). MMFF+LDA can improve the generalization and robustness so that it could be applied on different face data set. The weighted parallel fusion strategy is also proposed here(WMMFF+LDA). Experimental results on ORL, Yale, NUST603 and FERET human face data set respectively illustrate that the performance of the proposed method is more discriminant and its generalization and robustness are improved obviously too.Color and contour are also the important features of an image. A new color model used to detect flame in an image is found. The method to extract the contour feature of a flame image is also developed based on threshold of flame area.According to the color distribution and dynamic edge contour of a flame image, flames in video sequences are detected using color information and comparability of the edge of fire in image sequences. Many experiments illustrate that the system is able to work well and get high detection rate with a low false positive rate.