Research On Sketch Based Image Retrieval With Samples

This research is mainly to explore the algorithm of sketch retrieval related problems under few samples.In the field of deep learning,fewsample refers to the small number of samples used in the experiment.The purpose of few samples research is to enable the model to train and learn from very few samples,and finally achieve better results on the test data.Sketch-Based Image Retrieval(SBIR)refers to the retrieval of images from natural images database given a hand-drawn sketch.At present,compared with the readily available color images,the number of sketches is small and most of them are in different styles,lacking a lot of detailed information.Due to the different styles of sketches,many algorithms fail to fully extract the features of sketches.In order to solve this problem,this study proposes three data enhancement methods:eliminating blank pixels,sketch segmentation and combination,and generative adversarial networks.These three methods can enhance the data information of the sketch while eliminating the interference of the sketch style.In the field of sketch retrieval,there are large domain differences between sketches and color images,how to overcome the domain differences has always been the goal of researchers.This study proposes an algorithm combining prototypical networks and generative adversarial networks.The algorithm adopts the idea that the cluster points of categories are close to each other to increase the distance of the same category in different fields.At the same time,it uses the generative adversarial network to enhance the sketch features and generate difficult samples to further improve the robustness of the algorithm.In real-world scenarios,sketch retrieval often uses data that the algorithm model has never encountered,which is essentially a small-sample scenario.How to improve the accuracy of sketch retrieval in small sample scenarios has always been a hot topic.This study proposes an algorithm based on attention mechanism and learning metrics.This algorithm can make feature fusion have high feature migration ability and achieve optimal performance in multiple research scenarios.The contributions of this study are as follows:1)Three algorithms for enhancing the information of sketch data are proposed.2)The proposed prototype-based and generative adversarial network retrieval algorithms reduce inter-domain differences.3)Based on the attention mechanism and the learning measurement algorithm,the model has a high migration ability and has achieved the best performance in multiple research experiments.