YUV空间中基于稀疏自动编码器的无监督特征学习

Unsupervised Feature Learning with Sparse Autoencoders in YUV Space

现有无监督特征学习算法通常在RGB色彩空间进行特征提取,而图像和视频压缩编码标准则广泛采用YUV色彩空间。为了利用人类视觉特性和避免色彩空间转换所消耗的计算量,该文提出一种基于稀疏自动编码器在YUV色彩空间进行无监督特征学习的方法。首先在YUV空间随机采集图像子块并进行白化处理,然后利用稀疏自动编码器进行无监督局部特征学习。在预处理阶段,针对YUV空间亮度和色度通道相互独立的特性,提出一种将亮度和色度进行分离的白化措施。最后用学习到的局部特征在大尺寸图像上进行卷积操作从而获得全局特征,并送入图像分类系统进行性能测试。实验结果表明:只要对亮度分量进行适当的白化处理,在YUV空间中的无监督特征学习就能够获得相当于甚至优于RGB空间的彩色图像分类性能。

Existing unsupervised feature learning algorithms usually extract features in RGB color space, but YUV color space is widely adopted in image and video compression standards. In order to take advantage of human visual characteristics and avoid the calculation consumption caused by color space conversion, an unsupervised feature learning approach in YUV space based on sparse autoencoders is presented. First, image patches in YUV space are randomly sampled and whitened, and then are fed into sparse autoencoders to learn local features in an unsupervised way. Considering the characteristic that the luminance channel and chrominance channels are independent in YUV space, a whitening method which treats the luminance and chrominance separately is proposed in the pre-processing step. Finally, features learned over local image patches are convolved with large-size images in order to get global feature activations. Global features are then sent into image classification systems for performance testing. Experimental results reveal that unsupervised feature learning in YUV space achieves similar or even slightly better performance in color image classification compared with that in RGB space as long as the luminance component is whitened properly.