Research On Feature Extraction Method Of Joint Image

Feature extraction is one of the key issues in the field of pattern recognition, whose main task is to extract the most favorable information for pattern classification from the original sample images. In biometric identification, coding based method, due to its advantages, such as high accuracy, robustness, compact and fast matching, has been successful and widely used. This paper focuses on the FKP pattern recognition methods base on coding. But generally coding based feature extraction methods usually take into account the local information of the image while ignore the global information, local and global information fusion is one of the contents of this paper. The main contents of this paper are divided into four parts:(1) High order steerable filters are first employed to extract the continuous orientation feature map, then we use multilevel histogram thresholding method to quantize the feature map adaptively and the discrete orientations are used for coding a FKP image. Furthermore, we measure the similarity between two coded FKP images by designing an effective angular matching function. Experimental results on the PolyU FKP database demonstrated the accuracy of the proposed method was better.(2) We proposed a simple yet efficient and effective resolution approach called Binary Gabor Pattern. First, we convolve the given image with J Gabor filters, which share the same parameters except the parameter of orientation, to obtain even and odd features respectively. Then we can get J bits at each location by binarizing the responses. J bits can be assigned a unique integer, then we can extract the even and odd BGP features from given image and fuse the two features. The final classification is based on the image’s histogram of its BGP. Experiments conducted on Poly U FKP database demonstrated that our method works well whether in recognition rate or storage space.(3) Research on local and global information combination(LGIC) based FKP recognition method. First extract and match the local and global information separately, then the local and global are fused. The results confirmed that the local and global fusion method achieves higher accuracy.(4) Sparse multi-scale competitive code(SMCC) FKP recognition method was proposed. First sparse representation of FKP images are derived froml8 sDoGs, then encode the extracted feature matrix image using competitive coding rule. The results demonstrated that SMCC is superior to the other three kinds of coding methods based on the orientation information only.