Selective Deep Convolutional Neural Network for Low Cost Distorted Image Classification

Minho Ha, Younghoon Byeon, Youngjoo Lee, Sunggu LeeIEEE Access 2018

TL;DR

This paper introduces a "selective" deep convolutional neural network (DCNN) architecture designed for efficient classification of distorted images, a common challenge in real-world applications. By employing a two-stage approach with a tiny CNN to detect distortions and then routing to specialized, smaller DCNNs, the method achieves comparable accuracy to larger models while significantly reducing computational cost and energy consumption.

What Problem Does This Paper Address?

Deep convolutional neural networks (DCNNs) have revolutionized image classification, but their performance often degrades significantly when faced with distorted images (e.g., due to noise, blurring, or occlusion). This is a critical issue for real-world applications, especially in resource-constrained environments like mobile and embedded systems, where DCNNs are increasingly deployed. These systems demand low computational cost and energy efficiency, which are at odds with the large parameter counts and multiply-accumulate (MAC) operations typical of state-of-the-art DCNNs. Existing methods for improving distorted image classification often involve training with distorted data, which can be computationally expensive and may not generalize well, or simply reducing model complexity, which can lead to accuracy loss. The paper aims to bridge this gap by proposing a method that maintains high accuracy on distorted images while drastically reducing computational cost.

Methods

The core innovation is the "selective DCNN" architecture, which operates in two stages:

  1. Distortion Detection (Tiny CNN): A small, lightweight CNN is trained to distinguish between clean and distorted images. This tiny CNN is designed to be highly accurate for this specific task and has a negligible impact on overall computational cost.
  2. Specialized Classification (Dedicated DCNNs): Based on the output of the tiny CNN, the input image is directed to one of two dedicated DCNNs. These dedicated DCNNs are essentially "thinner" versions of established baseline DCNNs (like VGG16 or ResNet50), meaning they have fewer input/output channels. This reduction in channels directly leads to fewer weight parameters and MAC operations, thus lowering computational cost and energy usage. The paper uses versions of VGG16 and ResNet50 with channel depths reduced to 75%, 50%, and 25% of the original.

The authors also explore combining this selective architecture with established cost-reduction techniques like weight pruning and quantization to further enhance efficiency.

Datasets

The experiments were conducted using the CIFAR-100 dataset. A modified version of this dataset was created, consisting of a clean dataset and distorted datasets with three types of distortions:

  • Gaussian noise: Modeled with a standard deviation of 0.25.
  • Gaussian blurring: Modeled with a standard deviation of 1.00.
  • Occlusion: Modeled as a 16x16 block.

For experimental convenience, a distilled dataset was created and used repeatedly.

Key Insights and Ideas

  • Distortion Detection as a Pre-processing Step: The paper highlights that a simple, accurate distortion detection mechanism can effectively guide classification, allowing for specialized, more efficient models to be used for different image types.
  • "Thinner" DCNNs for Distorted Images: Instead of using a single, large DCNN for all cases, the selective approach leverages smaller, optimized DCNNs for distorted inputs, achieving significant cost savings without substantial accuracy loss.
  • Synergy with Existing Techniques: The selective DCNN architecture shows promise when combined with weight pruning and quantization, further amplifying the cost reduction benefits.
  • Trade-off Management: The approach explicitly acknowledges and manages the trade-off between classification accuracy and computational cost, allowing for a small accuracy drop (under 1%) in exchange for significant efficiency gains, which is crucial for mobile/embedded applications.

Key Results

  • Significant Cost Reduction: The selective DCNN architectures (e.g., 0.50-VGG16, 0.25-ResNet50) achieved substantial reductions in weight parameters and MAC operations compared to their baseline counterparts. For instance, selective 0.50-VGG16 reduced weights by approximately 73% and MACs by about 74% compared to 1.00-VGG16.
  • Comparable Accuracy: Despite the cost reduction, the selective DCNNs maintained classification accuracy comparable to the baseline models, often within 1% for clean images and for specific distortion types. For example, selective 0.75-VGG16 achieved 68.38% accuracy on Gaussian blurring, compared to 67.53% for the baseline.
  • Data Access Energy Savings: Data access energy per image was significantly reduced. For Gaussian noise with selective 0.50-ResNet50, energy was reduced by about 3.7 times. For Gaussian blurring/occlusion with selective 0.75-VGG16, energy was reduced by about 1.7 times.
  • Combined Techniques: When combined with weight pruning (80%) and quantization, the selective DCNN achieved similar accuracy with only about 5.6% of the data access energy of the baseline architecture.

Limitations and Future Work

  • Tiny CNN Performance: While the paper states the tiny CNN achieves near-perfect classification of clean vs. distorted images, its performance under extreme distortion levels or novel distortion types is not deeply explored.
  • Accelerator Design: The current energy estimation is based on the Eyeriss architecture. The authors note that for mobile/embedded systems processing images one by one, a dedicated accelerator design tailored to the selective DCNN might be more suitable and is a direction for future work.
  • Real-world Distortions: The experiments used specific, controlled distortions. Further validation on more complex, real-world distorted image datasets would be beneficial.

Summary updated

Abstract

Deep convolutional neural networks have proven to be well suited for image classification applications. However, if there is distortion in the image, the classification accuracy can be significantly degraded, even with state-of-the-art neural networks. The accuracy cannot be significantly improved by simply training with distorted images. Instead, this paper proposes a multiple neural network topology referred to as a selective deep convolutional neural network. By modifying existing state-of-the-art neural networks in the proposed manner, it is shown that a similar level of classification accuracy can be achieved, but at a significantly lower cost. The cost reduction is obtained primarily through the use of fewer weight parameters. Using fewer weights reduces the number of multiply-accumulate operations and also reduces the energy required for data accesses. Finally, it is shown that the effectiveness of the proposed selective deep convolutional neural network can be further improved by combining it with previously proposed network cost reduction methods.
BibTeX
@article{ha2018selective,
  author = {Minho Ha and Younghoon Byeon and Youngjoo Lee and Sunggu Lee},
  title = {{Selective Deep Convolutional Neural Network for Low Cost Distorted Image Classification}},
  journal = {IEEE Access},
  year = {2018},
}
Figure