Abstract: For the issue of the large image feature dimension in existing Content-Based Image Retrieval（CBIR） methods, a CBIR method based on improved Convolutional Neural Network（CNN） and bilinear model is proposed. Firstly, a low-dimensional pooling method is used to replace the pooling process in the traditional CNN to reduce the dimension of the image features. Then, two feature extractors are used for feature extraction based on the idea of bilinear model, and these two features are inner product at each image location to form the final image descriptor. Finally, the similarity is evaluated by calculating the Manhattan distance between images, so as to obtain the related images and their sorting. Experimental results show that the proposed method can accurately retrieve the relevant images, and has lower retrieval time and memory consumption.

Key words: Content-Based Image Retrieval（CBIR）, Convolutional Neural Network（CNN）, bilinear model, low-dimensional image representation, Manhattan distance

摘要： 针对现有基于内容的图像检索（Content-Based Image Retrieval，CBIR）方法中图像特征维度较大等问题，提出一种结合改进卷积神经网络（Convolutional Neural Network，CNN）和双线性模型的CBIR方法。采用一种低维度池化方法代替传统CNN中的池化过程，以此降低图像特征映射的维度。基于双线性模型的思想，使用两个特征提取器进行特征提取，并在每个图像位置上对两个特征进行内积，以形成最终的图像描述符。通过计算图像间的曼哈顿距离度量来评估相似性，获得相关图像及其排序。实验结果表明，该方法能够准确检索出相关图像，并具有较低的检索时间和内存消耗。

关键词: 基于内容的图像检索（CBIR）, 卷积神经网络（CNN）, 双线性模型, 低维度图像表示, 曼哈顿距离