Research On Image Feature Extraction Method Based On Multi-task Learning Technology

When the number of labeled training samples is very small, the sample information people can use would be very little and the recognition rates of traditional image recognition methods are not satisfactory. However, there is often some related information contained in other databases during extracting the features. The information contained in these databases is helpful to the feature extraction. Therefore we can use the useful information contained in these databases to help extract the features more effectively and then the recognition rates can be improved.In our paper we look back and learn the history and principles of multi-task learning. Then succeed to apply multi-task learning to feature extraction of images. According to the level of transferring knowledge in multi-task learning our researches are mainly divided into two parts, and they are feature extraction method based on transferring the projection transformation and feature extraction method based on transferring the samples. The feature extraction method based on transferring the projection transformation includes feature extraction based on the orthogonal projection vector and feature extraction based on the orthogonal sample set. The feature extraction method based on transferring the samples section is also divided into two parts. One uses the idea of multi-task learning and splits the single task into multiple tasks for classification. The other employs the idea of transferring the samples and constructs the training sample set of the target domain using the different information between the samples of the source domain. Feature extraction methods based on transferring the samples are only helpful to similar databases, but their performance is relatively stable. However, feature extraction methods based on transferring the projection transformation can transfer the knowledge from unlike databases.Our experiments results on the public AR, FERET, CAS-PEAL and Palmdata databases demonstrate that the proposed approaches are more effective than the general related feature extraction methods in classification performance.