Introduction: Image searching is a continual challenge even with the
many image retrieval models that have sprung up. Sketch-Based
Image Retrieval (SBIR) models attempt to solve this challenge by
searching using sketching. The existing SBIR algorithms have limited
performance because of ambiguities and variations in hand-drawn
sketches.
Aims: The aim of this work was to review and identify the strengths
and weaknesses of the existing SBIR models.
Materials and Methods: Articles were selected from Google Scholar
assessing strictly sketch construction models. Search terms include
sketch construction, sketch-based image retrieval, hypermedia,
multimedia, design strategies, and algorithms.
Results: The search returned 455 articles of which only 134 studies
met the inclusion criteria. 30 papers were on Convolutional Neural
Network (CNN) and hybrids. 6 on Contour and Stroke Segments. 4
on Generative Adversarial Network while 3 papers were on Deep
Hashing. 6 papers reported use of 3D-CNN-based methods while 85
papers used other methods like sparse coding and bag of regions.
Accuracy, recall and precision ranged from 59.47% to 99.4%,
20.10% to 47.70% and 33.40% to 51.00% respectively.
Conclusion: There are some promising SBIR models but lots of
effort is required if computational SBIRs are to be adopted. Most
studies did not include any performance metric which makes it
difficult to assess the performances of the algorithms proposed.
Researchers are advised to always report the performance
algorithms. The future plan is to develop a robust SBIR algorithm
which will accommodate handwriting ambiguity variations