Research On Object Localization In Large-scale Image Set

With the popularity of all kinds of image acquisition devices, image resources are increasing dramatically. How to find and locate a kind of object quickly in a large-scale image set is an important technical problem that needs to be addressed in practical applications. Object localization has alway been the hot spot in the field of multimedia technology. The image resources in practical application tend to have a large quantity and multiple categories. There has been the lack of efficient solutions.In recent years, there has been a lot of related work about object localization. Mainstream technologies are based on the sliding window or the Hough transform method. However, those technologies cause many inefficient operations and need to be improved in positioning effect when meeting large-scale and multi-label image set. This paper proposed a technical idea that involving the image classification technology into the object location processing. Acoording to the specific application scenarios in the object localization process, a new bag generation method is proposed in the multi-instance multi-label (MIML) learning framework and the image classification effect is improved; two methods to fuse the classification information into the object localization are proposed, both of them have achieved relatively good experimental results.The following work has been done in this paper:1. For the object localization task on the large-scale and multi-label image set, the idea that using the image classification information to improve the positioning effect was proposed. Image classification and object localization contain different interpretations of the image, their information are independent and they can benefit from each other. We provide efficient solutions for the object localization task on the large-scale and multi-label image set through involving the image classification information.2. For the image classification problem in the object localization tasks, the efficient MIML learning algorithm was used. And a new bag generator was proposed based on knowing the regional tag information in the training image set. Thus MIML algorithms can learn more useful information for learning, and the accuracy of the image classification was improved. Experimental results under multipl MIML algorithm show that:Multi-bow bag can bring better classification effect than the traditional SBN bag.3. In order to improve the performance of the object localization technology, this paper researched the approach of fusing image classification into object localization. According to the object localization task in the large-scaled image set containing multi-class objects, this paper proposed an effective scheme that a fast image classification was carried out before a precise object localization. And aiming at a higher precision in the object localization task, the paper proposed a new scoring mechanism of the optimal box which include the global classified information. The corresponding experimental results show that the former method shortens the processing time while obtaining a good average positioning accuracy; the latter method brings an improvement in the localization performance on many categories as it makes the scoring mechanism of the optimal box better.On the basis of the above-mentioned technical research, the paper design a-n image prototype system. The user can set the object classes that need to co-nsider and the image libraries that need to browser. Finally, the system provide an object hierarchy browser.