Sketch image recognition using deep features

Sketch is a special group of images, and the ability to recognize sketches is of great importance for many applications, including the human-computer interaction and childhood education. Previous approaches often pose this as a sub-class of general image classification which can be solved with a conventional pattern recognition method. In this thesis, instead of applying general image classification methods directly to sketches, a model using a new deep neural network that considers more of the unique characteristics of sketches has been developed and studied. Experiment results on the challenging sketch datasets demonstrate the superior performance of this present model in comparison to previous state-of-the-art methods. In addition to a novel method for sketch image classification, the topic has been further expanded to enable the matching of face sketches to photo images. This is of critical significance, especially in the law enforcement area to identify criminals. Previous work has attempted to address this task by exploring invariant features or looking for a shared subspace. In this paper, an end-to-end method has been proposed whereby the similarity score can be obtained directly when inputting a pair of sketch and photo-face images. In particular, this study investigates matching image pairs with more diversity than previous studies, considering such features as the presence of a beard, or haircut style, etc. The approach taken in this study is one utilizing a CNN based model, which is more robust and applicable to the complexities existing in the real world. The method presented exhibits good performance with both forensic sketch dataset and viewed sketches. Furthermore, in this study, the largest face sketch dataset has been created which can accommodate, in total, 14750 positive pairs and 985512 negative pairs for evaluation, which will facilitate future research.

Author dc:creator

• Jiang, Yuchao

Advisor dc:contributor.advisor

• Shen, Chunhua

• sketch
• CNN
• classification
• face recognition

DOI

10.4225/55/5b10b6edb84c6

OAI identifier oai:identifier

oai:digital.library.adelaide.edu.au:2440/112588