摘要
传统的SIFT向量图像检索法精度偏低、实用性较差,无法满足使用者的心理需求,为此提出一种基于尺寸不变特征转换向量的图像检索精度优化方法。采用倒排文件的图像检索方式,通过改良建树方法、检索方式和匹配度计算的优化,在满足实时性要求的前提下实现检索精度的提高;为SIFT特征向量构建新的聚类机制,以K均值聚类与分类相结合的方式对传统K均值聚类法进行改进,检索过程中利用得到的欧氏距离信息统一处理SIFT向量,进而简化向量间距的计算过程;基于改进单位化的处理方式有效控制大数据所引起的误差。测试结果表明,所提出的算法强化了树模型中节点的差异性,有效解决了按数量均分训练集数据而不是根据距离分配以及直接设定vocabulary tree层数的问题,大幅提高了检索的效率和精度。
The traditional SIFT vector image retrieval method has low accuracy and poor practicability,so it is hard to meet the psychological needs of users.Therefore,an image retrieval method based on inverted file was proposed to improve the retrieval accuracy by optimizing the tree building method,retrieval method and matching degree calculation on the premise of meeting the real-time requirements;a new clustering mechanism was constructed for SIFT feature vector;the traditional K-means clustering method was improved by combining K-means clustering and classification.The obtained Euclidean distance information was used to process the SIFT vector in a unified way,and the calculation process of vector spacing was simplified.The error caused by big data was effectively controlled based on the improved unit processing method.The test results showed that the proposed algorithm enhanced the differences of nodes in the tree model,effectively solved the problem of evenly dividing the training set data according to the number rather than assigning according to the distance and directly setting the number of layers of vocabulary tree and greatly improved the efficiency and accuracy of retrieval.
作者
朴林
PIAO Lin(Liaoning Teachers College for Nationalities,Shenyang,Liaoning 110032,China)
出处
《河北北方学院学报（自然科学版）》
2022年第3期26-31,共6页
Journal of Hebei North University:Natural Science Edition
关键词
尺度不变特征
图像检索
精度优化
向量
scale invariant feature
image retrieval
precision optimization
vector
