PaperlayerView exact revision on arXiv

Source-linked AI summary

ColorNet: Investigating the importance of color spaces for image classification

Shreyank N Gowda, Chun Yuan

arXiv:1902.00267v1cs.CV

TL;DR

The paper examines whether relying directly on RGB limits image classification and studies color-space transformations as alternative representations. It combines predictions from multiple color spaces using small networks, reporting higher accuracy and strong results with substantially fewer parameters than larger DenseNet models. The approach also has preprocessing and computation overhead, especially for larger images.

  • Problem

    Most recent image-classification models use RGB inputs directly, while the paper investigates whether alternative color spaces better represent some image classes and can reduce parameter requirements.

  • Method

    The model converts each RGB image into multiple color spaces, processes them with small Densenets, and combines their outputs.

  • Results

    The approach reports higher accuracy from combining color spaces, comparable accuracy to a 25.6M-parameter DenseNet with 1.75M parameters on CIFAR datasets, and new state-of-the-art results on four datasets.

  • Takeaways & Limitations

    Different image classes can be better represented in different color spaces, making combined color-space predictions useful for accurate classification with smaller models.

  • Takeaways & Limitations

    For larger images such as ImageNet, color-space preprocessing and multiple Densenets create computational overhead despite using smaller and wider Densenets.

Abstract

from arXiv · show

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.

1 Introduction

The paper argues that image classification should treat color-space transformations as distinct image representations rather than relying only on RGB inputs. It proposes combining predictions from multiple color spaces to improve accuracy while using relatively small models.

  • Motivation: Most recent classification models use RGB images directly without color-space transformation.The paper frames color spaces as alternative numerical representations of the same image.
  • Color-space representations: Color spaces organize colors as numerical representations, including RGB, HSV, XYZ, LAB, transmission spaces, LCH, and CMYK.The paper describes these spaces as differing in channel meanings and transformation properties.
  • Motivation: The paper treats transformations into different color spaces as new images from the computer’s perspective.This motivates testing whether alternative representations change classification behavior.
  • Proposed approach: The proposed approach classifies images in different color spaces with small networks and combines their final-layer outputs.The method seeks outputs that are less correlated so the combined prediction can use complementary information.
  • Contributions: Certain image classes are better represented in particular color spaces, and combining color-space outputs gives higher accuracy than using each space individually.These are presented as core findings of the paper’s introduction.
  • Contributions: A 1.75M-parameter model achieves accuracy similar to a 25.6M-parameter DenseNet model on CIFAR datasets and reports new state-of-the-art results on CIFAR-10, CIFAR-100, SVHN, and ImageNet.The paper positions this as evidence that smaller multi-color-space models can remain highly accurate.

2 Related Works

Prior work improved image classification through deeper and densely connected networks, but often used many parameters. Earlier color-space studies produced differing results, motivating combined color-space modeling.

  • Deep architectures: Deep convolutional networks improved image classification, while greater depth introduced vanishing-gradient difficulties.Normalization, intermediate layers, identity connections, and short paths were explored to support deeper networks.
  • Deep architectures: DenseNet connected every layer to every other layer and achieved strong image-classification results.The approach was designed to improve information flow and address vanishing-gradient problems.
  • Parameter efficiency: Recent high-performing approaches often required many parameters, motivating efforts to preserve accuracy while reducing model size.The paper frames parameter reduction as a central design objective.
  • Color-space preprocessing: Earlier studies used color-space preprocessing for tasks including skin detection, pixel classification, and soccer-image analysis.Reported results differed across applications and color spaces.
  • Color-space preprocessing: The paper builds on conflicting earlier findings by combining color-space models to seek higher classification accuracy.The proposed direction treats differing color-space results as complementary rather than selecting one universally best space.

3 Proposed Approach

The proposed approach evaluates color-space transformations and combines predictions from models trained on different representations. It uses late fusion with small, wide DenseNets to combine complementary class-level information while limiting parameter growth.

  • Color-space analysis: Different color spaces produced different class-level accuracies, including 82 percent versus 72 percent for CIFAR-10 class 4.The comparison used confusion matrices across RGB, HSV, YUV, LAB, HED, LCH, XYZ, and YPbPr.
  • Color-space analysis: The analysis found that color-space outputs were not perfectly correlated and that different spaces represented some classes better.These observations provided the basis for combining color spaces.
  • Fusion strategy: Late fusion was selected because early fusion reduced parameters but also produced lower accuracy.The chosen approach combines model outputs rather than concatenating transformed inputs at the start.
  • Color-space analysis: 78.89 percent increased to 86.14 percent when outputs from seven color-space CNNs were combined.This improvement motivated using smaller DenseNets for each color space.
  • Model architecture: The model uses seven DenseNets, each receiving a different color-space representation, followed by a dense layer producing weighted class predictions.The selected spaces are RGB, LAB, HSV, YUV, YCbCr, HED, and YIQ.
  • Model architecture: The input RGB image is converted simultaneously into six additional color spaces before all seven representations are classified.The final dense-layer output serves as the classification score.
  • Trade-offs: Multiple color conversions add time overhead, while combining seven DenseNets can substantially increase model complexity.The approach therefore faces a computation and parameter trade-off despite using smaller sub-networks.

4 Experimental Analysis

The evaluation compares ColorNet with recent image-classification approaches across CIFAR-10, CIFAR-100, Imagenet, and SVHN. ColorNet reports state-of-the-art results while using compact models, including a 1.75M-parameter model that matches or exceeds substantially larger DenseNets on CIFAR datasets.

  • Evaluation setup: The experiments evaluate ColorNet on CIFAR-10, CIFAR-100, SVHN, and Imagenet using comparisons with recent state-of-the-art approaches.The comparisons include Network in Network, All-CNN, Highway networks, FractalNet, ResNet, Wide-ResNet, and DenseNet where specified.
  • CIFAR-10: ColorNet-40-48 achieves a 1.54 error rate on augmented CIFAR-10, reported as a new state-of-the-art classification result.The result is reported for CIFAR-10 with data augmentation.
  • CIFAR-10: A 1.75M-parameter ColorNet-40-12 exceeds DenseNet-BC-250-24 with 15.3M parameters and nearly matches DenseNet-BC-190-40 with 25.6M parameters on CIFAR-10.This comparison accompanies the augmented CIFAR-10 result.
  • Imagenet: ColorNet-121, which replaces the proposed model’s subnetworks with DenseNet-121 networks, obtains a new state-of-the-art accuracy on Imagenet.Imagenet comparisons report Top-1 and Top-5 accuracy and single-crop versus 10-crop error rates.
  • SVHN: ColorNet-40-48 obtains a 1.12 error rate on SVHN, while ColorNet-40-12 with 1.75M parameters exceeds DenseNet-BC-250-24 with 15.3M parameters.The 1.12 error rate is reported as a new state-of-the-art classification result.

5 Conclusion

The study finds that image classes are represented differently across color spaces, motivating a combined model that uses seven selected spaces. This approach achieves strong accuracy with a small parameter budget, but preprocessing and multi-network computation create overhead for larger images.

  • Color-space analysis: Different color spaces represent image classes differently, even when overall accuracy changes little.Confusion-matrix analysis showed that classification results were not perfectly correlated across color spaces.
  • Proposed model: The combined model uses seven color spaces: RGB, YIQ, LAB, HSV, YUV, YCbCr and HED.LCH, YPbPr and XYZ were discarded after reducing overall model accuracy.
  • Proposed model: Seven DenseNet branches transform the input into different color spaces and combine their outputs through a dense layer for weighted predictions.The architecture uses a separate DenseNet for each transformed input.
  • Results and limitations: 1.75M parameters matched models with 27M parameters while reaching state-of-the-art results.The approach uses smaller and wider DenseNets to reduce computation and hyperparameter overhead.
  • Results and limitations: Preprocessing and multiple DenseNet branches impose time and computation overhead, especially for larger ImageNet images.For CIFAR and SVHN, preprocessing can be performed in real time; for larger images, it creates cost overhead.
Loading 1902.00267v1…