Research On Efficient Coding Of Natural Images And Ensemble Transfer Learning Approaches

Image categorization plays an important role in Computer Vision andImage Understanding. There exist two key problems in discriminativeclassification models, namely “representation” and “learning”. On onehand, better feature representation will greatly benefit later classificationsegment. Efficient coding of natural images performs well as it hastheoretical support in biological vision. On the other hand, traditionalmachine learning techniques can’t obtain reliable classifier when trainingset is small or the distribution of training and test data are not identical.Transfer learning could share knowledge among different but relativedomains or tasks, leveraging existing knowledge for current tasks.This thesis includes the following contents:(1) Common image feature representations and their differences incomputation process are introduced, within which feature representationbased on efficient coding theory is specifically focused on.(2) An image bases learning and coding approach in independentsubspaces is proposed. Image bases are learned via Independent SubspaceAnalysis in an unsupervised manner. Regularly divided patches are codedand pooled using Spatial Pyramid Matching model.(3) Summarizes the research status and relative technologies oftransfer learning. Transfer learning algorithms under the framework ofensemble learning are reviewed and analyzed in detail.(4) A modified ensemble transfer learning algorithm based onmulti-source TrAdaboost is proposed. Covariate shift is incorporated intothe loss function and transferability for source domains is evaluated toavoid negative transfer and promote recognition accuracy.