A Feasibility Study on Local Hand-crafted Feature Descriptors for Sketch-based Image Retrieval

Image retrieval plays a major role in medical diagnosis, forensic labs, military, crime prevention, and web searching. Most of the image retrieval systems are text-based, but images normally have little or not carry any textual information. The sketch-based Image Retrieval (SBIR) method allows the user to search natural images using freehand sketches instead of text. From the previous investigation, it is found that SBIR may cause the unmatching of the sketches with the database set due to the user's sketches or the algorithm itself. This work is to study the effectiveness of three local handcrafted descriptors which are Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), and Oriented FAST and Brief Descriptor (ORB) for SBIR. It is done by comparing the similarity score between the sketch image and the real image using three different local descriptors. The results demonstrate that each used descriptor produces different matched keypoints and the feature vectors for the similarity measure. To calculate the similarity percentage, Euclidean distance was chosen among the other distance measurement methods. From the results obtained, SIFT has the highest percentage followed by SURF and ORB.