Research On Cross-scenario Image Automatic Annotation Method Based On Multi-instance Learning

With the rapid development of mobile Internet, e-commerce has already deeplychanged our traditional lifestyle. High-performance intelligent terminal also makesit possible to take photos anywhere and anytime, and thus generating a huge volumeof cross-scenario images. How to retrieve and recognize the objects in thosecross-scenario images, containing complex background and irregular patterns, hasbecome a hot resrarch topic. Providing textual description of an image, the tag playsan important role in the field of image processing. However, it is impossible tomanually tag all these massive, complex cross-scenario images. Therefore, theapproaches to automatically annotate these cross-scenario images are needed whichcan then be used for other image processing applications.Image annotation is one of the most important problems in image processingareas. The traditional algorithms extract either the global or local features to projectimages onto a new feature space, and then annotate images through general machinelearning algorithms. These approaches seldom consider the relationship betweendifferent objects and tags, and can not solve the background difference problem.Therefore, this paper proposed the cross-scenario image annotation approach basedon the multi-instance learning which extends the application scope of the traditionalannotation algorithms. And this paper mainly completes the following tasks:First of all, according to the application requirements, this paper proposed animage annotation approach to solve the recognition problem of cross-scenarioimages. Considering the complexity of the background and the richness of thecontent, we design a special pre-process and feature extraction method with respectto the cloth image;Then, considering the relationship between the objects and tags, we consider tocombine the multi-instance learning framework and the image annotation together.We also design the corresponding automatic annotation procedure;Finally, according to the image annotation method proposed in this paper, wechoose the multi-instance learning as our image classification algorithm to label animage. To consider image semantic ambiguity, we further expend the problem as amulti instance multi label learning problem. We then modify the existingMIMLSVM method by integrating a new bag embedding method. Finally, weintroduce the semi-supervised MIML. The effectiveness of the proposed approachesis shown in the experiments.