The Effect of Color Channel Representations on the Transferability of Convolutional Neural Networks

Abstract

Image classification is one of the most important tasks in computer vision, since it can be used to retrieve, store, organize, and analyze digital images. In recent years, deep learning convolutional neural networks have been successfully used to classify images surpassing previous state of the art performances. Moreover, using transfer learning techniques, very complex models have been successfully utilized for other tasks different from the original task for which they were trained for. Here, the influence of the color representation of the input images was tested when using a transfer learning technique in three different well-known convolutional models. The experimental results showed that color representation in the CIE-L*a*b* color space gave reasonably good results compared to the RGB color format originally used during training. These results support the idea that the features learned can be transferred to new models with images using different color channels such as the CIE-L*a*b* space, and opens up new research questions as to the transferability of image representation in convolutional neural networks.