摘要
针对传统基于哈希算法的图像检索方法生成的哈希编码难以保留图像语义相似性,从而导致检索准确率较低的问题,文中结合深度神经网络和哈希算法的优点,使用两次特征二值化对图像的哈希编码进行优化。对于不同查询图像基于第一次特征二值化进行阈值调整,然后根据调整后的阈值完成第二次特征二值化并利用优化后的哈希编码完成相似图像的检索。采用公开的图像数据库Cifar-10及MNIST得到的实验结果显示,所用方法得到的图像识别准确率高于传统方法以及采用固定阈值进行特征二值化的相关方法,证实了所提出方法的有效性。
The hash code generated by image retrieval methods based on traditional hash had difficulty in pre-serving semantic similarity among images, which led to low retrieval accuracy. Combined with the advantages of deepneural network and hash algorithm, the hash code of images was optimized by using two teature binarization. For theditterent query images, the threshold was adjusted based on the first teature binarization, and then the second teaturebinarization was completed according to the threshold adjusted above. Finally, the optimized hash code was utilized toperform the retrieval of similar images. The results obtained from the public Cifar - 10 and MNIST database showedthe image recognition accuracy given by the proposed method was higher than those obtained from the traditionalmethod or other related methods using fixed thresholds for teatures binarization, which confimmd the ettectiveness ofthe proposed method.
作者
王伟栋
李菲菲
谢林
陈虬
WANG Weidong, LI Feifei, XIE Lin, CHEN Qiu(School of Optical - Electrical and Computer Engineering, University of Shanghai tor Science and Technology, Shanghai 200093, China)
出处
《电子科技》
2018年第10期48-52,共5页
Electronic Science and Technology
基金
上海高校特聘教授(东方学者)岗位计划(ES2012XX
ES2014XX)
关键词
深度神经网络
哈希算法
图像检索
特征二值化
阈值调整
deep neural networks
hashing algorithm
image retrieval
teature binarization
threshold adjustment