Research On Part-based Object Recognition And Its Applications

Object recognition in still images is a basic task in computer vision. Object recognition, which provides object-related information, such as categories and positions of objects, has long been a critical technology that bridges low-level image processing tasks and high-level tasks. Multiple class object detection is an approach to implement object recognition. Object detectors annotate objects in images with bounding boxes as while as obtain categories of objects, they thus can provide compact object-related information.Though research on object detection has been developed vigorously in re-cent years, there are still some problems: 1. There may be different ways for the detection of objects in different categories; Also because of the variance of objects within a category, the similarity of objects in different categories, and the complexity of backgrounds, it may be difficult to build a unified model for all categories. 2. Object detection has long been considered as a binary clas-sification problem, but this formulation ignores orders of examples, especially pair-wise and list-wise information. 3.Because of the binary classification for-mulation of object detection, research on object description has long been con-strained on word labels, and can not obtain more compact description.To address the problems mentioned above, this paper models structures of objects to get more powerful object detection models, formulates object detec-tion problem as a ranking problem to mine more information of datasets, and generates more compact description of objects. The main contributions in this paper are:(1)This paper proposes RDPM(Ranking Deformable Part Model). Based on the modeling of the structure of objects in DPM(Deformable Part Model),this paper introduces a ranking loss function, which is proved to be a GC-CCP(Generalized Concave-Convex Programming) problem, and proposes an optimization algorithm of the loss function. Experiments on open dataset show that ranking models have better detection performance than original models.(2) This paper proposes an object ranking framework for object detection,and applies the framework on k-best results from DPM. This paper uses an ensemble of multiple rankers to generate a ranking algorithm that is suitable for object detection, models are trained on dynamic datasets to obtain a good performance. Experiments show that this framework not only promotes the performance of DPM, but also achieves the state-of-the-art results.(3) This paper proposes an image description generation model centered on part-based object detection systems. There are three types of information in part-based object detection results: categories of objects, positions of objects,layouts of parts. The description generation model first builds alignment rela-tionships between these types of information and language units: mapping cat-egories to nouns, positions to positional adverbs and layouts of parts to posture verbs. After the relationships are built, this paper proposes a description gener-ation algorithm covering sentence aggregation and surface realization. Experi-ments show that compared with description methods using deep learning tech-nologies, the method in this paper can generate multiple sentences, and more compact descriptions about object positions and object postures. Based on the description generation model, this paper implements a demo system for image description generation centered on part-based object detection, and a demo sys-tem for image retrieval.