基于卷积神经网络和流形排序的图像检索算法

Image retrieval algorithm based on convolutional neural network and manifold ranking

针对基于内容的图像检索(CBIR)中低层视觉特征与用户对图像理解的高层语义不一致以及传统的距离度量方式难以真实反映图像之间相似程度等问题,提出了一种基于卷积神经网络(CNN)和流形排序的图像检索算法。首先,将图像输入CNN,通过多层神经网络对图像的监督学习,提取网络中全连接层的图像特征;其次,对图像特征进行归一化处理,然后用高效流形排序(EMR)算法对查询图像所返回的结果进行排序;最后,根据排序的结果返回最相似的图像。在corel数据集上,深度图像特征比基于场景描述的图像特征的平均查准率(m AP)提高了53.74%,流形排序比余弦距离度量方式的m AP提高了18.34%。实验结果表明,所提算法能够有效地提高图像检索的准确率。

In Content-Based Image Retrieval（ CBIR）, the low-level visual features are not consistent with the high-level semantic features captured by human, and it is difficult to reflect the similarity of images by traditional distance measurements.To solve these problems, an image retrieval algorithm based on Convolutional Neural Network（ CNN） and manifold ranking was proposed. Firstly, the image dataset was put into CNN, image features were extracted through the fully connected layers of the network after supervised learning; secondly, the image features were normalized and then Efficient Manifold Ranking（ EMR） algorithm was used to return the ranked scores for query images; finally, the most similar images were returned to users according to the scores. In corel dataset, the mean Average Precision（ m AP） of deep image feature was 53. 74% higher than that of the scene descriptor features, and the m AP of efficient manifold ranking was 18. 34% higher than that of the cosine distance. The experimental results show that the proposed algorithm can effectively improve the accuracy of image retrieval.