Sketch-based Image Retrieval Method

Abstract

As the sketch describing the peripheral contour information of the object is more aligned with people's subjective visual experience of the object in the process of perception.Therefore,sketch-based image retrieval technology did a lot of attention. In the sketch retrieval system, the system will process the sketches such as edge extraction, contour tracking, feature point assignment, feature conversion, and imilarity matching, and finally provide users with several target images with higher similarity. In the process of searching for the target image, the main basis for users to search is the shape information of the image. Therefore, in the sketch retrieval system, only the contour feature of the image is considered, and the color feature and texture feature of the image are not considered. This paper proposes a sketch image retrieval method based on imageassisted crossdomain learning, which effectively establishes the crossdomain matching relationship between sketches and images, and improves the effectiveness of features and the accuracy of matching. Considering the lack of sketch data, this paper uses largescale image data for pretraining and crossdomain initialization of the network. In order to reduce the modal difference between the natural image and the hand drawn sketch, the Gb contour extraction method is used to extract the contour of the image, remove internal useless lines, and improve the modal similarity between the contour and the sketch. In addition, considering that the key information of sketch image retrieval is shape, on the basis of the original classification loss function of the network model, a shape regression loss item is introduced, and joint training optimization is performed to fully tap the shape consistency between the sketch and the image relationship.