dc.description.abstract | As touching devices have rapidly proliferated, sketch has gained much popularity as an
alternative input to text descriptions and speeches. This is due to the fact that sketch
has the advantage of being informative and convenient, which have stimulated sketchrelated
research in areas such as sketch recognition, sketch segmentation, sketch-based
image retrieval, and photo-to-sketch synthesis. Though these eld has been well touched,
existing sketch works still su er from aligning the sketch and photo domains, resulting
in unsatisfactory quality for both ne-grained retrieval and synthesis between sketch and
photo modalities. To address these problems, in this thesis, we proposed a series novel
works on free-hand sketch related tasks and throw out helpful insights to help future
research.
Sketch conveys ne-grained information, making ne-grained sketch-based image retrieval
one of the most important topics for sketch research. The basic solution for this task
is learning to exploit the informativeness of sketches and link it to other modalities.
Apart from the informativeness of sketches, semantic information is also important to
understanding sketch modality and link it with other related modalities. In this thesis,
we indicate that semantic information can e ectively ll the domain gap between sketch
and photo modalities as a bridge. Based on this observation, we proposed an attributeaware
deep framework to exploit attribute information to aid ne-grained SBIR. Text
descriptions are considered as another semantic alternative to attributes, and at the same
time, with the advantage of more
exible and natural, which are exploited in our proposed
deep multi-task framework. The experimental study has shown that the semantic
attribute information can improve the ne-grained SBIR performance in a large margin.
Sketch also has its unique feature like containing temporal information. In sketch synthesis
task, the understandings from both semantic meanings behind sketches and sketching
i
process are required. The semantic meaning of sketches has been well explored in the
sketch recognition, and sketch retrieval challenges. However, the sketching process has
somehow been ignored, even though the sketching process is also very important for us
to understand the sketch modality, especially considering the unique temporal characteristics
of sketches. in this thesis, we proposed the rst deep photo-to-sketch synthesis
framework, which has provided good performance on sketch synthesis task, as shown in
the experiment section.
Generalisability is an important criterion to judge whether the existing methods are able
to be applied to the real world scenario, especially considering the di culties and costly
expense of collecting sketches and pairwise annotation. We thus proposed a generalised
ne-grained SBIR framework. In detail, we follow the meta-learning strategy, and train
a hyper-network to generate instance-level classi cation weights for the latter matching
network. The e ectiveness of the proposed method has been validated by the extensive
experimental results. | en_US |