hello-Sketch: An attempt of image retrieval based on sketch
This topic, 'sketch based image retrieval', is my subject that I work on in the scientific research training that I participated in during my sophomore year(presumably from Janurary 2019 to December 2019). In this work, I try to design a hand-craft local feature that can be effective for image classification and retrieval task.
My work can be summarized by following steps (for one RGB image, it may experiecnce following operations):
- get grayed
- get edge extracted
- edge fitted with straight line
- a series local feature (say with shape [m,1]) extracted with a kind of feature design to compose the image's global feature (say with shape [m,n])
- the image's glocbal feature is feed into a fisher vector to be encoded to a one-dimension vector (say with shape [a, 1])
every RGB image in dataset should go through all 1~5 process
every sketch image in dataset should go through 2~5 process
while all image in dataset have got their feature vector, companied with their labels, following tasks can be done
- for image classification task, feed the feature vactors and labels to train a svm model.
- for image retrieval task, the similarity of two images can be caculated using euclidean distance or other distance norm, according to the distance, retrival task can be fullfilled.
dataset used:extended ETHZ shapes