Research On Multi-label Printing Image Retrieval Based On Deep Learning

With the popular use of smartphone cameras,the amount of image data has been rapidly increasing.Printing pattern retrieval is a special case in image retrieval and has great potential in commercial applications and industrial production.Meanwhile,printing pattern is diversiform and subjective,so it is challenging to retrieval the pattern efficiently in a large-scale database.Recently,many CBIR(Content Based Image Retrieval)methods have been proposed with the development of deep learning.However,most of these methods have adopted binary supervision indicating whether a pair of images are of the same class or not,which is not suitable for printing pattern retrieval.Therefore,this thesis focuses on the novel printing pattern retrieval and apply multi-labels to image retrieval tasks inveterately.1.A graph convolutional network based on attention-driven Graph Convolutional Network Based on Attention Mechanism(GCN-AM)is proposed to extract representative features of multi-label printing patterns.Firstly,this method uses the attention module to activate the region of interest corresponding to the label.Secondly,it puts the learned label embedding from multi-semantic attention module into the graph convolutional network(GCN)to capture the dependency of labels,so GCN-AM can extract high-level semantic features with label semantics and spatial relationships for retrieval.Experiments on the two printing pattern data sets constructed in this thesis demonstrate that the attention area of our method is more meaningful and discriminative under multistage training.We can see that the proposed method highlights the relevant semantic regions of labels through visualization,and it achieves higher retrieval accuracy than existing deep learning methods in dense or sparse multi-label printing patterns.2.A multi-label deep hash algorithm with ranking loss embedding is proposed with the combination of deep network and hash coding technology.The algorithm considers not only the fine-grained relationship between multi-label images,but also images distribution in the hash space.Additionally,it uses quantization loss to control the quality of hash coding.Experiments on two multi-label printing pattern data sets show that the proposed method outperforms the competing methods in multiple indicators.It further solves the problem of information loss caused by encoding and reduces the retrieval time greatly.3.The two proposed multi-label deep retrieval algorithms and some existing image retrieval algorithms are applied to the multi-label print pattern data set constructed in this thesis.The thesis verifies the effectiveness of the proposed algorithms,analyzing their advantages and disadvantages from many aspects.The results show that the algorithms proposed in this thesis achieve higher retrieval accuracy,lower retrieval complexity and better scalability.