Sketch_based Image Retrieval Using Convolutional Neural Network with Multi_step Training

Abstract

The expansion of touch-screen devices has provided the possibility of human-machine interactions in the form of free-hand drawings. In sketch-based image retrieval (SBIR) systems, the query image is a simple binary design that represents the mental image of a person with the rough shape of an object. A simple sketch is convenient and efficient for recording ideas visually, and can outdo hundreds of words. The objective is to retrieve a natural image with the same label as the query sketch. This article presents a multi-step training method. Regression functions are used in the deep network structure to improve system performance, and various loss functions are employed for a better convergence of the retrieval system. The convolutional neural network used has two branches, one related to the sketch and the other related to the image, and these two branches can have the same or different architecture. After four training steps, a 56.48% MAP was achieved, indicating the desirable performance of the network.