Sketch Based Image Retrieval Using Deep Feature

In recent years,owing to the popularity of touch-screen devices and handwriting devices as well as the popularity of expression pack,a lot of people began to try to use sketches to replace the text or language to abstract expression.In the process of drawing sketches,people can easily express anything and ideas in the mind,and then find the similar image content in the active image database by drawing the sketch.This way of using hand-painted sketches to express ideas conforms to the most primitive perception of mankind.And,people with different levels of knowledge,or different culture,or different levels of painting,or different ages can use this way to express the natural images they saw.Therefore,sketch-based image retrieval has also been researched by a lot of scholars at home and abroad.Most of the current hand-drawn sketches are based on manual features,including binary features and natural image features.And then the middle-level feature code is used to synthesize the final sketch feature representation.However,sketch has great abstraction,and the same objects drawn by different people have a large appearance difference,which leads to the fact that the performance of sketch-based image retrieval based on the traditional hand-feature representation is too dependent on the drawing level of the hand drawing lines.Therefore,those retrieval system based on the traditional manual characteristics performs not good and is lack of robustness.In this paper,we show that deep convolutional neural networks(DCNN)is also suitable for sketch-based image recognition,i.e.,using sketch as query to retrieve natural images in a large dataset.In order to solve this kind of cross-domain problem,we propose to train CNN jointly using image data and sketch data in a novel way,and then training to get the deep sketch feature(DSF).The learned deep feature is very effective for sketch-based image retrieval.Using simple Euclidean distance on the learned feature can significantly outperform the previous state-of-the-arts.In addition,we found that in the standard sketch classification benchmark,training DCNN with pre-training and a feasible data argumentation can largely surpass human-level performance.