Research On Image Classification Based On Label-embedding

Standard supervised learning requires that all training samples are labeled,but this method is often limited by limited labeled training data.Zero-shot learning is a classification problem without labeled training samples.The emergence of network as an information sharing tool makes the machine has more data to learn.However,although an almost infinite number of images can be obtained through a media sharing site such as Flick,there are still insufficient labeled images to build a model for identifying a large number of visual target categories.Therefore,in recent years,zero-shot learning began widespread concern.The research of zero-shot learning is of great significance for increasing data size and category.This paper studies image feature extraction,classifier and side information.The image feature plays a decisive role in the final classification accuracy.In recent years,convolutional neural network has achieved great success in the image recognition community.Convolutional neural network serving as a supervised learning,is a good solution to the traditional image classification problem,but also for the zero-shot learning problem to provide the deep image features.Due to the good semantic characteristics of deep features,improving the classification accuracy substantially.Traditional zero-shot learning classification method based on multiple linear regression,project visual features into semantic embedding space.In this paper,the spatial distribution of the visual features,especially the deep features,is investigated.The learning target recognition classification algorithm based on inverse projection is proposed.Experiments on three data sets show that the classification accuracy of zero-shot is improved significantly.Zero-shot learning classification method uses semantic embedded space as the bridge of knowledge transfer,the most commonly used attribute vector.However,attribute vectors still require a lot of artificial definitions,thus limiting the scalability of the zero-shot learning method.So we introduce the word vector generated by word2 vec method as side information to replace the traditional attribute vector in order to improve the scalability of the algorithm.