Concerning the high time-consuming and low recognition accuracy of Bag of Feature (BOF) model, a new improved BOF model was proposed to improve the accuracy and efficiency of target recognition, and it was also applied to cow recognition. The optimized Histogram of Oriented Gradient (HOG) feature was introduced to feature extraction and description of the images; then the Spatial Pyramid Matching (SPM) principle was used to generate the histogram representation of images based on visual dictionary; finally, the histogram intersection kernel defined in this paper was used as the kernel function of the classifier. The experimental results on the data set in this paper (including 15 kinds of cows with 7500 images of cow heads) showed that the recognition rate of the algorithm was improved by an average of 2 percentage points by using the BOF model based on SPM; compared with Gauss kernel, the recognition rate of the algorithm was increased by an average of 2.5 percentage points by using the histogram intersection kernel; compared with traditional HOG feature, the recognition rate of the algorithm was improved by an average of 21.3 percentage points by using optimized HOG feature, and the computation efficiency of the algorithm was improved by an average of 1.68 times; compared with Scale Invariant Feature Transform (SIFT) feature, the computation efficiency of the algorithm was improved by an average of nearly 7.10 times as well as ensuring the average recognition accuracy reached 95.3%. Analysis results indicate that this algorithm has good robustness and practicability in cow individual recognition.

Image interpolation is an important method to magnify images with low resolution to adapt to the target screens. To preserve the geometry feature of the original image is an effective way to improve the quality of magnified images. This paper proposed a new method to interpolate images based on Doo Sabin subdivision. The method adopted the essential idea of subdividing the quadrilateral mesh to enhance the sampling images of low resolution. Firstly, part of the data of high resolution images was obtained by mapping low resolution images. Secondly we classified the unknown pixels of high resolution images according to their geometric features. Then we interpolated all the unknown pixels by the assigned pixels. Values of the unknown pixels were the weighted average of their neighboring pixels. The weighted strategy was deduced by Doo Sabin subdivision. Experiments show that our method can preserve the sharp feature of image edges, decrease zigzags and achieve better results than the previous methods.