Research On Key Technologies Of Scene Classification And Object Recognition

With the scalable development of image data, the demand of automatical imageunderstanding becomes more and more urgent. As the important topic of imageunderstanding, the image categorization has attracted great attention in a wide range ofrelevant computer vision applications, such as content-based image indexing andretrieval (CBIR), remote sensing image processing, intelligent video surveillance, andintelligent vehicle/robot navigation. Thus far, the image categorization has emerged asone of the most attractive areas in the field of image understanding and computer vision.Although considerable progress has been made, image categorization remains achallenging issue, due to the complexity of object imaging.In this thesis, the methods of image categorization are investigated. Promisingresults are attained in several aspects including the local image descriptors and theapplication of descriptors to image categorization, the low-level modeling basedapproach for scene classification, local feature based approach for image categorizationand the saliency object detection. The main work and achievements of the thesis can beconcluded as follows:(1) Approaches for local image descriptors.In order to represent local image features more effectively for scene classificationand object recognition, two new visual descriptors called GBPWHGO (Gradient BinaryPattern and Weighted Histogram of Gradient Orientation) and GLID (Gradient LocalInvariant Descriptor) are introduced. The GBPWHGO descriptor is mainly used forscene classification, which focuses on the distinctiveness of local feature description.For the GLID descriptor, we are most concerned with the issue of invariance. By usingthe GLID descriptor, we want to compensate for various affine deformations in objectrecognition, such as scale changes, rotations and changes of viewpoint.(2) Low-level modeling based approach for scene classification.Most of the current scene classification approaches can be roughly divided into twobroad families: low-level and semantic modeling based approaches. Although thelow-level modeling based approaches offer simplicity and low computational cost, theyare normally applied to classify images into a small number of scene categories, andoften exhibit poor performance. The semantic modeling based approaches can deal witha large number of scene categories and present inspiring performance, but theseapproaches usually bring high computational cost and memory consumption. In thisthesis, we propose an alternative solution to figure out the above-mentioned problems.Our method largely follows the low-level modeling based strategy. However, it retainsthe advantages of both the low-level and semantic modeling strategies, while at the same time getting over the above-mentioned weaknesses of these two strategies.(3) Local feature based approach for image categorization.Recently, the most popular local feature based image categorization approaches arebag of features (BoF) method and spatial pyramid matching (SPM) method. The BoFmethod represents an image as an orderless collection of local features, whichdisregards all information about the spatial layout of the features, affecting the accuracyof recognition. The SPM method partitioned an image into increasingly fine sub-regionsand computed histograms of local features inside each sub-region, which incorporatespatial information to a certain extent. In this thesis, we propose an alternative solutionto improve the SPM method in local feature extraction and the incorporation of spatialinformation. Experiment results demonstrate that our approach produces commendableperformances.(4) Saliency object detectionIn this thesis, we propose a global color contrast based saliency object detectionalgorithm. The proposed algorithm is simple, yields full resolution saliency maps, anddetects saliency object automatically. The saliency object detection has proven to besuccessful in various computer vision problems such as image categorization, objectdetection, and image segmentation.