• Qianqian Gu

Student thesis: Phd


This project presents the idea of constructing a Computational Aesthetics Learning System in the Traditional Chinese Painting(TCP) domain. The system tends to imitate the human visual system of processing TCP data and related aesthetic information. To achieve this target, four individual tasks are set. They are feature extraction and analysis, image classification, object detection, and neural style transfer in the TCP domain. Four components were designed to consist of our computational aesthetic system. First, an image representation descriptor combines both hand-crafted and deep learning representations by employing Hash mapping with Hamming distance and similarity thresholding. This descriptor interprets TCP domain knowledge by applying Sobel's kernel to HOG, and TCP colour palette based on Basic Color Terms [7]. Second, a set of TCP style classifiers, both hand-crafted ML (SVM) and DL (VGG) models, are introduced to distinguish TCP styles and content schools by using the image representations resulting from the stage-one descriptor. The classification process allows us to verify which models and which features are more sensitive to TCP so that we can integrate them to train the backbone network for later use. Next, a two-stage Deep CNN object detector is proposed to improve our aesthetic learning system's ability to recognise objects of TCP artworks. Objects of art are the basis of aesthetic analysis. The Assembled Region Proposal Network (A-RPN), which we promoted, enhanced a general RPN by a CN-Kitten RPN with transposed convolutional feature information. The RoI pooling layer is replaced by a top rank-voting model and position-sensitive score maps. This model is sensitive in detecting small objects and incorporating translation variance. The A-RPN significant outperformed YOLO 2 and original Faster R-CNN with p-values of as p
Date of Award1 Aug 2022
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorDavid Morris (Supervisor) & Xiaojun Zeng (Supervisor)


  • Feature Extraction
  • Traditional Chinese Painting
  • Neural Style Transfer
  • Machine Learning
  • Computational Aesthetics
  • Object Detection

Cite this