[Paper Review] ColorNet: Investigating the importance of color spaces for image classification

Shreyank N Gowda, Chun Yuan|arXiv (Cornell University)|Feb 1, 2019

Brain Tumor Detection and Classification22 citations

TL;DR

ColorNet proposes a novel image classification framework that simultaneously processes a single RGB image through seven distinct color spaces—RGB, HSV, LAB, YUV, YCbCr, YIQ, and HED—using separate small, wide DenseNets for each. By fusing predictions from these color-space-specific branches, the model achieves state-of-the-art accuracy on CIFAR-10, CIFAR-100, SVHN, and ImageNet with only 1.75 million parameters, outperforming larger models like DenseNet-BC-190-40 (25.6M parameters).

ABSTRACT

Image classification is a fundamental application in computer vision. Recently, deeper networks and highly connected networks have shown state of the art performance for image classification tasks. Most datasets these days consist of a finite number of color images. These color images are taken as input in the form of RGB images and classification is done without modifying them. We explore the importance of color spaces and show that color spaces (essentially transformations of original RGB images) can significantly affect classification accuracy. Further, we show that certain classes of images are better represented in particular color spaces and for a dataset with a highly varying number of classes such as CIFAR and Imagenet, using a model that considers multiple color spaces within the same model gives excellent levels of accuracy. Also, we show that such a model, where the input is preprocessed into multiple color spaces simultaneously, needs far fewer parameters to obtain high accuracy for classification. For example, our model with 1.75M parameters significantly outperforms DenseNet 100-12 that has 12M parameters and gives results comparable to Densenet-BC-190-40 that has 25.6M parameters for classification of four competitive image classification datasets namely: CIFAR-10, CIFAR-100, SVHN and Imagenet. Our model essentially takes an RGB image as input, simultaneously converts the image into 7 different color spaces and uses these as inputs to individual densenets. We use small and wide densenets to reduce computation overhead and number of hyperparameters required. We obtain significant improvement on current state of the art results on these datasets as well.

Motivation & Objective

• To investigate whether different color spaces significantly impact image classification performance.
• To determine if certain color spaces are more suitable for specific image classes or datasets.
• To reduce model complexity while maintaining or improving accuracy by leveraging multi-color-space input.
• To explore whether early color space transformation before feature learning improves representation learning.
• To develop a lightweight yet highly accurate model by combining multiple color-space representations in a single architecture.

Proposed method

• The model takes a single RGB image as input and simultaneously transforms it into seven distinct color spaces: RGB, HSV, LAB, YUV, YCbCr, YIQ, and HED.
• Each color-space representation is fed into a separate, small, and wide DenseNet (DenseNet-40-12) to extract features independently.
• The outputs from all seven DenseNets are concatenated and passed through a final fully connected layer for weighted fusion and final classification.
• The use of small and wide DenseNets reduces the number of parameters and hyperparameters while mitigating vanishing gradient and overfitting issues.
• Color space transformations are applied using standard mathematical formulas, including linearization of sRGB and conversion to CIE XYZ and LAB via matrix multiplication.
• The final prediction is obtained by averaging or learning-weighted fusion of the individual branch outputs.

Experimental results

Research questions

• RQ1 How does the choice of color space affect image classification accuracy across diverse datasets?
• RQ2 Can combining multiple color-space representations improve classification performance compared to using RGB alone?
• RQ3 Are there specific color spaces that are more effective for certain image classes or dataset characteristics?
• RQ4 To what extent can model size be reduced while maintaining or improving accuracy through multi-color-space input?
• RQ5 Does early color space transformation lead to better feature learning than processing raw RGB images?

Key findings

• ColorNet-121 achieved a top-1 error rate of 17.65% and top-5 error rate of 5.22% on ImageNet, setting a new state of the art at the time of publication.
• Colornet-40-12 with only 1.75 million parameters outperformed DenseNet-BC-250-24 (15.3M parameters) on SVHN, achieving a 1.59% error rate.
• Colornet-40-48 achieved a 1.12% error rate on SVHN, representing a new state of the art for that dataset.
• The model significantly outperformed larger models: for example, Colornet-40-12 surpassed DenseNet-BC-190-40 (25.6M parameters) on CIFAR-100 with fewer parameters.
• The confusion matrix analysis revealed that certain classes were better classified in specific color spaces, justifying the multi-color-space fusion strategy.
• Color spaces such as LCH, YPbPr, and XYZ were found to degrade performance and were therefore excluded from the final model.