Image Topic Identification by Machine Learning COS511 Project Report
چکیده
Object recognition and image categorization are two related topics in computer vision and image processing, both being very challenging. In existing researches, however, these two topics do not overlap much. The available methods for object recognition is rather domain specific, and researchers are more likely to focus on distinguishing fine grained differences between objects of the same class, like human face recognition, hand written digit recognition and so on. Often, a specific class of objects has its own model, which often do not generalize well. On the other hand, image categorization takes a more coarse way, which is carried out in a whole image, and does not work well when a single image is made up of different components. Recent researches have started to use region based methods to do image categorization[3], but the topic is still to be well exploited. In this project, we go along that direction, which is somewhere in between of object recognition and image categorization. We bring up the problem of topic identification and try to tackle the problem by leveraging machine learning methods. Often, images contain different topics, or objects. By saying topics, we emphasize the semantic meaning of the object, or the class of object, but care less about detailed difference between objects of the same class. Basically, the topics of the image are just what are included in the image. For example, a picture of Furnace Greek Inn may have the following topics:inn, trees, mountains and sky. The main problem to be solve in this project is: given an image, identify the topics it contains. We use machine learning methods to construct models from available annotated image set. We compare whole image based approach and segment based approach, believing that segment based topic recognition will be more accurate. Two machine learning algorithms, naive Bayes classifier and artificial neural networks are used, with there performance compared in this specific learning problem. We will explained the image segmentation and feature extraction methods in Section 2, and discuss the machine learning methods in Section 3. In Section 4, we incorporate the machine learning method within a framework of segment based learning. We show experimental results in Section 5, and the report is concluded in Section 6.