Application Of Feature Extraction Based On Image Segmentation To Zero-Shot Learning

As an important information carrier, the digital image has been widely used in many fields. Almost all the information of the digital image hides in the low-level features. Feature extraction is an important stage for the image processing. Whether successfully extracting the image features or not is a key to distinguish or connect images. The zero-shot learning technology has widely used the unsupervised feature extraction, which make the computer successfully simulate the cognitive habits of human beings. With zero-shot learning as the application background, to solve the problem of high dimension of extracted texture feature, and not take into account the overall characteristics and local characteristics at the same time, as well as the extraction of color features to ignore the relationship between color and space, in order to get better performance in zero-shot learning, The main research works are described as follows:This thesis proposes a local threshold-based watershed algorithm. To solve image sampling problem in the convolution process of Gabor template and image, the lack of flexibility and incomplete information of the regional information of the fixed window sampling and uniform sampling, and in order to fully extracts the image details. Firstly, extracting the outline of the image using Sobel operator. Then because the image to be segmented is complex, the Bernsen algorithm obtaining several local thresholds of image to make it binaryzation. Finally, the binaryzation image is segmented by watershed algorithm.A Gabor based on LTW texture feature-extraction algorithm is proposed. The algorithm is used to extract the texture features of the image. Firstly, there isn't a common method for setting parameters of the Gabor filter, so a common method for setting parameters of the Gabor filter is proposed in this thesis. The whole feature of segmented image blocks is extracted as the local feature of the whole image. Then, by combining the proposed setting parameters method and the LTW algorithm, compute the mean and variance of the overall texture as the final texture feature. Therefore, the proposed LTW-Gabor algorithm can significantly reduce the dimension of the texture feature. Finally, the extracted texture features are applied to the zero-shot learning.A feature extraction algorithm based on GRW-Lch is proposed. Firstly, to solve the phenomenon of ignoring the relationship between color and space in the color histogram, the gradient of the color image is reconstructed. Then, extract the color feature on segmented image blocks with the color histogram using local color histogram. Finally, the extracted color features are applied to the zero-shot learning.Make experiments on three different data sets, i.e, Public Figure Face(Pubfig), Outdoor Scene Recognition(OSR) and Attribute Discovery(shoes) datasets. The comparison experiments on several datasets show that the proposed method can obtain higher accuracy of attribute prediction and zero-shot learning recognition rate compared with the features of the traditional zero-shot learning.