Convolutional autoencoder-based color image classification using chroma subsampling in YCbCr space

We propose a convolutional autoencoder neural network for image classification in YCbCr color space to reduce computational complexity. We first learned local image features from image patches in YCbCr space with a sparse autoencoder and then convolved them with large images to obtain global features. Chrominance components were subsampled before convolution as it is permitted to reduce bandwidth for chrominance components in YCbCr space. We then adopted an algorithm to resize the convolved features in chrominance components by shifting the elements after convolution. Global features were finally fed into a softmax classifier to test the classification accuracy. Experimental results reveal that the convolutional neural network in YCbCr space is able to obtain a reduction of at least 21.6% in time consumption compared to the RGB representation with a slight loss in accuracy.