Le deep learning pour la classification des images dans différents système de couleur.

Abstract

Image classification is fundamental in the field of artificial intelligence, recently Deep Residual Learning [1] and the new coming model CapsulNet[2] have shown state-of-the-art performance for image classification tasks , they takes data-sets mostly as input in the form of RGB images even though we have many other colorspaces available. In this thesis we try to understand the impact of image color-space on the performance of CNN models in Image classification . We evaluate this on CIFAR10 [3]data-set by converting it into five other color-spaces HLS, HSV, LUV, LAB, YUV and trained each one of them in two different deep learning architecture models namely ResNet20 and CapsulNet , the results obtained show a minor change in accuracy but even a small percentage may make the difference, in the other hand LUV is good alternative it show improvement about 0.92% comparison to RGB in ResNet20 and 0.39% in CapsulNet.