Image retrieval method based on sketch

Abstract

The invention discloses an image retrieval method based on a sketch. According to the method, domain migration learning from an image domain to a sketch domain is realized, and a network can output effective features which are adaptive to the sketch domain and have the distinguishing performance. The method comprises the following steps of extracting a Canny edge of an image to serve as network training data; carrying out network pre-training based on millions of images of ImageNet; sending the Canny edge of the image to the network for training, and realizing domain migration learning; sending the sketch and the Canny edge of the image to the trained network, and extracting the features respectively; and computing cosine distances between the extracted features and sorting, and realizingK nearest neighbor retrieval. According to the method, the defects of sketch data of the network in a training process are fully compensated, the effect of sketch training is improved, a cross-domaindifference between the sketch and the image is reduced, and a good retrieval effect is achieved.

Description

technical field [0001] The invention relates to the technical fields of image retrieval and computer vision, in particular to an image retrieval method based on a sketch. Background technique [0002] With the rapid increase of Internet media image data, efficient and accurate image retrieval technology has become an urgent need. The early text-based retrieval methods have shortcomings such as relying on manual annotation and ambiguity, and cannot be applied to large-scale image retrieval. Therefore, content-based image retrieval has become a research hotspot. In content-based image retrieval, sketch-based retrieval can express user intention more conveniently and intuitively, and has received extensive attention. With the popularization of touch-screen devices, it is more and more convenient to draw hand-drawn sketches. Hand-drawn sketches can accurately express users' retrieval intentions. Therefore, image retrieval based on sketches has very important research significa...

AI Technical Summary

Problems solved by technology

This method is simple and effective, and has a fast retrieval speed, but it also has certain defects: the text description of the image needs manual annotation, and the Internet image data is growing rapidly, and the speed of manual annotation of images is much lower than the growth rate of the number of images , so this method is unsustainable; secondly, keywords often cannot accurately describe the user's search intention; thirdly, the manual tagging process will introduce the subjective judgment of the tagger, and different taggers have different perceptions of images, so There will be some ambiguity in the understanding of the image

However, the limitations of these handcrafted features affect the performance of sketch-based image retrieval

At the same time, the method based on deep learning is a data-driven method. The lack of sufficient training data can not make the method based on deep learning achieve good results, and it is easy to cause over-fitting problems.

Images