Research On Label Learning Methods Based On Deep Hashing

The task of Label Learning is to predict the label information for unknown samples.It is a significant task in artificial intelligence,widely used in Image Retrieval,Face Recognition,Image Segmentation,Object Tracking,and other fields.Many researchers have studied Label Learning and proposed many excellent algorithms.There are two main learning paradigms for Label Learning:(1)Mapping samples directly to the label space through a mapping function.Although this approach can well model Label Learning,these methods typically have very high computational complexity;(2)Hashing-based Label Learning.Hash technology is widely used in large-scale retrieval systems due to its advantages of low storage and high retrieval efficiency.Applying hash technology to Label Learning can increase the scalability of Label Learning.In addition,in the prediction stage,hashing-based Label Learning is obtained by predicting labels from known data.Therefore,the label information generated by hashingbased Label Learning is more consistent with the true data distribution than the traditional method of independently generating labels from models.In Label Learning,Label Distribution Learning(LDL),as a new learning paradigm,has achieved relatively high accuracy and efficiency.However,there are still some important issues that need to be addressed,leading to the existing hashing-based Label Distribution Learning methods only achieving suboptimal performance:(1)Shallow semantic modeling leads to the inability to effectively capture implicit semantics in the learning process;(2)Existing methods simply construct a matrix with marked similarity to preserve the semantic relationships of instances,which cannot completely model the internal semantic relationships of instances;(3)The existing methods have not well solved the problem of quantization errors in Hashing,which will lead to significant information loss.In response to these issues,this thesis proposes a Deep Discrete Hashing for Label Distribution Learning(DDH-LDL)method,which constructs the first deep learning framework based on hash for label distribution learning.Specifically,DDH-LDL captures implicit semantic information through multi-layer nonlinear transformations,while preserving the semantic relationships of instances into hash codes through the semantic information aggregation function of Graph Convolutional Networks.In addition,this thesis has carefully designed a discrete optimization module that is seamlessly integrated into the proposed deep hash framework to reduce binary quantization errors.Experiments on several widely tested datasets have verified the superiority of this method in terms of accuracy and efficiency.In addition,this thesis studies Image Set Labeling Learning.Image Set Labeling Learning has received widespread attention in recent years,with the task of learning the labeling information of image sets.Although current methods have high accuracy,there are still some unsolved problems:(1)There are certain similarities and complementarities between images within the set,but existing methods do not well consider the relationship between images in the image set;(2)There is no effective aggregation constraint within the set for each image within the set,so the internal relationship of the images within the set cannot be well modeled.To solve these problems,this thesis proposes Self-Attention Aggregation Hashing for Image Set Label Learning.This method is based on hash retrieval technology,which can quickly perform retrieval and predict unknown samples based on the label of the retrieval results,reducing the uncertainty of label learning,and making it more reliable.This method fully considers the relationships between various images within an image set,constrains the center of the set based on the label,and preserves the sample similarity relationship into the hash code.This method is more consistent with the unique characteristics of Image Set Label Learning.This thesis verifies the effectiveness of this method on multiple datasets.