PERCEPTUAL SHAPE FEATURE BASED IMAGE CODING FOR VISUAL CONTENT CLASSIFICATION AND OBJECT RECOGNITION

Abstract

The most essential technique in creating agents with the ability to process and understand the content of visual data is object recognition, which includes image content classi cation, and object localization. Deep convolutional neural networks' (CNNs) performance gain in computer vision, there still are application scenarios with limited training data and computing power for which using deep CNN based methods is not feasible. On the other hand, the human engineered image representations require less training data and computing power and can be enhanced by importing domain specif c knowledge. These representations may also bene fit from the human vision characteristics in reducing the gap between computed image representations and human vision perception. In this thesis we have proposed four methods to improve image classi cation and object localization. All these methods utilize the perceptual shape features of image since it is proved that the human vision perception on objects mostly relies on shape features of the objects, while color and texture are utilized as extra sources to complete this perception. In the rst method, we have created a static dictionary of perceptual shape features based on N-gram model and used that in combination with spatial pyramid matching to represent images. In the second method, a dynamic dictionary from image edge segments is formed where these segments are obtained from an octave of image in di erent scales. The third method considers the curve partitioning points as descriptive features of the image and created a dynamic dictionary from descriptors of these points. The proposed object localization method utilizes the perceptual shape features of the image to improve the location of objects determined by object recognition module. The initial location may be obtained by any object recognition method, then the proposed method iteratively merges the edge segments with the detected object using a best rst search strategy. These proposed methods have been evaluated on di erent benchmark image datasets. Judging on the overall performance of the proposed method, it is expected that the proposed methods would bring some useful alternatives to support e cient tool development for applications lacking training data or no training data at all.