结合多标签共现关系的深度哈希胸部X光影像检索

Deep Hash Retrieval of Chest X-ray Images with Multi-label Co-occurrence Relationship

为使哈希检索模型兼具强判别力和小量化误差,现有方法通常需要优化多个损失函数,导致模型训练存在困难.通过最大化连续哈希码与对应正交化二值码之间相似性,尽管只需优化单个损失函数就能实现这一目标,然而在多标签图像检索任务中,这类方法忽视了标签间的语义相关性,导致检索性能下降.本文提出一种基于标签共现的深度哈希检索算法.首先通过挖掘阳性疾病标签的共现信息,利用图卷积神经网络建模多标签共现关系,动态生成哈希目标.其次,通过引入标签平滑交叉熵损失函数,进一步增强图像哈希码与标签哈希目标的一致性.在胸片数据集上的实验结果表明,方法在关键性能指标上优于同类算法,验证了建模多标签共现关系对提升模型性能的重要性.

To enhance the discriminative power and minimize the quantization error of hash retrieval models,existing methods typically need to optimize multiple loss functions,which makes the training of the model difficult.By maximizing the similarity between continuous hash codes and the corresponding orthogonal binary codes,this goal can be achieved with a single loss function optimization.However,such methods neglect the semantic relevance between labels in multi-label image retrieval tasks,leading to degraded retrieval performance.This paper proposes a deep hashing retrieval algorithm based on label co-occurrence.Firstly,by mining the co-occurrence information of positive disease labels,a graph convolutional neural network is used to model the multi-label co-occurrence relationships,dynamically generating hash targets.Secondly,by introducing a label-smoothing cross-entropy loss function,the consistency between the image hash codes and label hash targets is further enhanced.Experimental results on a chest X-ray dataset show that the proposed method outperforms existing algorithms on key performance metrics,confirming the importance of modeling multi-label co-occurrence relationships in improving model performance.