Small Object Detection and Recognition Using Context and Representation Learning

Small object detection and recognition is very common in real world applications, such as remote sensing images analysis for Earth Vision, Unmanned Aerial Vehicle vision and video surveillance for identity recognition. Recently, the existing methods have achieved impressive results on large and medium objects. But the detection and recognition performance for small or even tiny objects is still far from satisfaction. The problem is highly challenging because small objects in low-resolution images may contain fewer than a hundred pixels, and lack sufficient details. Context plays an important role on small object detection and recognition. Aiming to boost the detection performance, we propose a novel discriminative learning and graph-cut framework to exploit the semantic information between targeting objects’ neighbours. What is more, to depict a local neighbourhood relationship, we introduce a pairwise constraint into a tiny face detector to improve the detection accuracy. At last, to describe such a constraint, we convert the problem of regression that estimates the similarity between different candidates into a classification problem that produces the score of classification for each pair of candidates. In representation learning, we propose an RL-GAN architecture, which enhances the discriminability of the low-resolution (LR) image representation, resulting in comparable classification performance with that conducted on high-resolution (HR) images. In addition, we propose a method based on a Residual Representation to generate a more effective representation of LR images. The Residual Representation is adapted to fuel back the lost details in the representation space of LR images. At last, we produce a new dataset WIDER-SHIP, which provides paired images of multiple resolutions of ships in satellite images and can be used to evaluate not only LR image classification but also LR object recognition. In the domain of a small sample training, we explore a novel data augmentation framework, which extends a training set to achieve a better coverage of varying orientations of objects in a testing data, so as to improve the performance of CNNs for object detection. Then, we design a principal-axis orientation descriptor based on super-pixel segmentation to represent the orientation of an object in an image. We propose a similarity measure method of two datasets based on a principal-axis orientation distribution. We evaluate the performance and show the effectivity of CNNs for object detection with and without rotating images in the testing set. Dissertation is directed by Professor Xiangjian He and DoctorWenjing Jia of University of Technology Sydney, Australia, and Professor Jiangbin Zheng of Northwestern Polytechnical University, China.

Author dc:creator

• Xi, Yue

Statement dc:rights

• The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
• au.edu.uts.lib/ppc
• info:eu-repo/semantics/openAccess

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/10453/148886

OAI identifier oai:identifier

oai:opus.lib.uts.edu.au:10453/148886