深度PCA子空间极限学习机图像检索算法

Image Retrieval Algorithm Based on Deep PCA Subspace Extreme Learning Machine

传统的基于内容的图像检索方法缺少自主学习能力,图像表达能力不强,严重制约其图像检索性能,而深度学习模型为图像检索提供了新思路.本文提出一种深度PCA子空间极限学习机图像检索算法.首先将图像进行分块处理,采用多层级联主成分分析作为卷积滤波层,将图像映射到深层PCA子空间,然后通过深度极限学习机获得深层子空间稀疏特征,实现图像的深层特征提取.最后对特征进行哈希编码,利用编码实现快速图像检索.在MNIST、CIFAR-10和CALTECH256等数据集上的实验结果表明,该算法在训练效果和训练时间上都具有较好的性能,与卷积神经网络等深度学习框架相比,具有结构简洁、收敛速度快等优点.

The traditional content-based image retrieval method lacks self-learning ability,and the image expression ability is not strong,which seriously restricts its image retrieval performance. The deep learning model provides a newidea for image retrieval. This paper proposes an image retrieval algorithm based on deep PCA subspace extreme learning machine. First,the image is divided into blocks,and multi-level cascaded principal component analysis is used as the convolution filter layer to map the image to the deep PCA subspace. Then,the deep subspace sparse feature is obtained by the depth extreme learning machine to realize the deep feature extraction of the image. Finally,the feature is encoded by hash and the image is retrieved quickly by coding. The experimental results on MNIST,CIFAR-10 and CALTECH256 showthat the proposed algorithm has better performance in training effect and training time.Compared with convolutional neural networks and other deep learning frameworks,the proposed algorithm has the advantages of simple structure,fast convergence,and so on.