基于多层次特征表示的图像场景分类算法

Image scene classification algorithm based on multi-level feature representation

传统场景分类采用底层尺度不变特征变换(SIFT)特征,运用词袋(BoW)模型以及空间金字塔(SPM)模型进行分类判别。然而,单一的低层描述的识别精度有限,无法有效表征内容多变的场景图像。本文提出基于多层次特征表示的图像场景分类算法,利用滑动窗均匀采样图像块,分别提取图像块的密集SIFT特征和卷积层卷积神经网络(CNN)特征,使用聚集局部描述符编码(VLAD)方法分别编码图像块的局部特征,将一幅图像的多个图像块特征顺序级联形成该幅图像的描述,由此构建包含局部语义信息的低层图像描述和中层图像描述。与此同时,将图像的低层描述与中层描述融合到图像的全连接层的高层语义中,从而获得整合了局部空间信息和全局语义信息的精确图像表示。本文在两个常用的场景数据集上进行了分类实验,结果表明,融合多层次特征描述的图像表示能够取得更好的分类结果。

The traditional scene classification uses the bag of words (BoW) model and the spatial pyramid matching (SPM) model with scale invariant feature transform (SIFT) features for classification discrimination. However, the single low-level description fails to represent the scene images due to the complexity and variability of scenes. This paper proposes an image scene classification algorithm based on multi-level feature representations. The convolutional neural networks (CNN) features from the convolutional layer and the dense SIFT features of the image blocks, sampled by sliding windows evenly, are extracted and encoded by the vector of locally aggregated descriptors (VLDA) method, respectively. The encoding SIFT features and CNN features of multi-blocks are sequentially cascaded respectively to form a low-level description and middle-level description. Both descriptions contain the local semantic information of the image. Meanwhile, the low-level description and the middle-level description of the image are integrated into the high-level semantic features of the full-connected layer of the image, so that a more accurate image representation is obtained by integrating the local spatial information and the global semantic information. In this paper, scene classification experiments are performed on two commonly used scene datasets. The experimental results show that the fusion representation of multi-level feature descriptions achieves better classification results.