Please use this identifier to cite or link to this item: https://scholarbank.nus.edu.sg/handle/10635/134430
DC FieldValue
dc.titleSCALE-ROBUST DEEP LEARNING FOR VISUAL RECOGNITION
dc.contributor.authorJIE ZEQUN
dc.date.accessioned2016-12-31T18:01:20Z
dc.date.available2016-12-31T18:01:20Z
dc.date.issued2016-08-17
dc.identifier.citationJIE ZEQUN (2016-08-17). SCALE-ROBUST DEEP LEARNING FOR VISUAL RECOGNITION. ScholarBank@NUS Repository.
dc.identifier.urihttp://scholarbank.nus.edu.sg/handle/10635/134430
dc.description.abstractIn recent years, deep learning has achieved great progress in almost all the visual recognition tasks. Nevertheless, deep learning lacks both image-level and object-level scale-robustness, making it difficult to handle the recognition tasks where testing images are in wide range of scales or contain objects with significantly diverse scales. In this thesis, we focus on improving both image-level and object-level scale-robustness for deep learning, leading to better recognition performance faced with the images and objects having large scale ranges. First, scene recognition requires scale invariance for better recognizing the captured images of diverse scales. To achieve scale invariance for scene recognition, we proposed a framework integrating the recent powerful deep convolutional networks and locality-constrained linear coding. Second, we proposed an end-to-end object detection framework based on fully convolutional networks (FCN) to detect vehicles and pedestrians. Third, existing localization strategies generally fail in producing satisfying localization accuracy for small objects. We thus proposed a novel scale-aware pixel-wise object proposal network to tackle the challenges. Fourth, in object detection, it is common that multiple objects are shown in one captured image. Existing localization algorithms usually search for possible object regions over multiple locations and scales separately, which ignore the interdependency among different objects. To incorporate global interdependency between objects into localization, we propose an effective Tree-structured Reinforcement Learning (Tree-RL) approach to sequentially search for objects by fully exploiting both the current observation and historical search paths.
dc.language.isoen
dc.subjectdeep learning, visual recognition
dc.typeThesis
dc.contributor.departmentMECHANICAL ENGINEERING
dc.contributor.supervisorLU WEN-FENG
dc.contributor.supervisorTAY ENG HOCK
dc.description.degreePh.D
dc.description.degreeconferredDOCTOR OF PHILOSOPHY
dc.identifier.isiutNOT_IN_WOS
Appears in Collections:Ph.D Theses (Open)

Show simple item record
Files in This Item:
File Description SizeFormatAccess SettingsVersion 
JieZ.pdf27.87 MBAdobe PDF

OPEN

NoneView/Download

Page view(s)

254
checked on Feb 2, 2023

Download(s)

2,090
checked on Feb 2, 2023

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.