PaperlayerView exact revision on arXiv

Source-linked AI summary

Some Improvements on Deep Convolutional Neural Network Based Image Classification

Andrew G. Howard

arXiv:1312.5402v1cs.CV

TL;DR

The paper addresses how to improve deep convolutional neural-network image classification using additions to an existing winning pipeline. It applies expanded training and test-time transformations, greedy prediction selection, and complementary higher-resolution models, achieving a 13.55% top-5 error rate in ILSVRC2013.

  • Problem

    The paper investigates improvements to state-of-the-art deep convolutional neural-network image classification systems evaluated in ILSVRC.

  • Method

    The system adds full-image crop sampling, color transformations, multiscale and multiview test predictions, greedy selection, and complementary higher-resolution models.

  • Results

    13.55% top-5 error improved on the previous year’s 16.4% top-5 error by 20% in relative terms.

  • Takeaways & Limitations

    The described transformations, greedy selection, and higher-resolution models provide additions intended to improve convolutional-neural-network image classification systems.

Abstract

from arXiv · show

We investigate multiple techniques to improve upon the current state of the art deep convolutional neural network based image classification pipeline. The techiques include adding more image transformations to training data, adding more transformations to generate additional predictions at test time and using complementary models applied to higher resolution images. This paper summarizes our entry in the Imagenet Large Scale Visual Recognition Challenge 2013. Our system achieved a top 5 classification error rate of 13.55% using no external data which is over a 20% relative improvement on the previous year's winner.

1 Introduction

The paper investigates additions to the 2012 winning convolutional-network system for ImageNet classification and summarizes the authors’ 2013 challenge entry. It reports a 13.55% top-5 error rate, improving on the previous year’s 16.4% result by 20% relatively.

  • ILSVRC evaluates state-of-the-art image classification using 1.2 million images across 1000 classes.
  • The authors extend the 2012 winning system with additional training and test-time transformations, prediction selection, and other model changes.The paper is organized around training transformations, test-time prediction transformations, and complementary higher-resolution models.
  • 13.55% top-5 error improved on the previous year’s 16.4% top-5 error by 20% in relative terms.Clarifai achieved 11.74% top-5 error in the 2013 competition.
  • The baseline model uses the 2012 winning system’s code and training methodology, while doubling fully connected layer sizes without improving top-5 error.

2 Additional Data Transformations for Training

The training pipeline expands image variation by sampling crops from the full image and applying additional color manipulations. These transformations increase available training examples and target translation, lighting, and color invariance.

  • These transformations generate training examples spanning image variations and help the network learn invariance to lighting and color properties.
  • Previous cropping discarded roughly 30% of image pixels, while full-image cropping incorporates those additional pixels.
  • Random 224x224 crops from 256xN or Nx256 images generate more training examples and learn broader translation invariance.Using the full image can include features such as a cat’s tail or both ears that square cropping may omit.
  • The method adds random contrast, brightness, and color manipulations to existing random lighting noise.The three manipulations are applied in random order with enhancement values between 0.5 and 1.5.

3 Additional Data Transformations for Testing

The testing pipeline combines multiple crops, flips, scales, and image views to improve predictions while controlling computational cost. A greedy algorithm selects compact prediction subsets with accuracy close to or better than using all transformations.

  • Multiple test-time transformations: Combining 5 translations, 2 flips, 3 scales, and 3 views produces 90 predictions per model.The three scales are 256, 228, and 284, and the three views cover different square regions of the image.
  • Multiple test-time transformations: Three image views incorporate all pixels from 256xN or Nx256 images by using left or upper, center, and right or lower square views.
  • Results: Table 1 compares new training and testing transforms with the previous baseline and shows that doubled fully connected layers do not substantively improve top-5 validation error.
  • Reducing predictions: The greedy approach reduces the delay caused by the expanded prediction set using a simple selection procedure.The 90-prediction combination is almost an order of magnitude larger than the previous 10 predictions.
  • Reducing predictions: The greedy algorithm starts with the best prediction and adds predictions until no further improvement remains.
  • Reducing predictions: The first 10 selected predictions are almost as accurate as all 90, while 15 predictions slightly improve on the 90-prediction baseline.

4 Higher Resolution Models

Higher-resolution models use the same architecture and can be initialized from trained low-resolution models. They provide complementary predictions, with a base-plus-high-resolution combination outperforming five base models for top-1 classification and matching them for top-5 classification.

  • Model motivation: Higher-resolution models retrain the network at larger image scales because the base network no longer performs well after excessive image scaling.
  • Model construction: The models use 224x224 crops from 448xN or Nx448 images so four corner crops do not overlap at test time.
  • Model construction: Initializing higher-resolution models with trained low-resolution models reduces training from 90 epochs to 30 epochs.
  • Test-time predictions: The higher-resolution model increases predictions per model to 162 using 9 crops, 2 flips, 3 scales, and 3 views.
  • Results: A base-model and high-resolution-model combination is better than five base models for top-1 classification and equivalent for top-5 classification.

5 Results

The submitted ILSVRC2013 system combined base and high-resolution neural networks, improving over the previous year’s winner while remaining behind Clarifai. Table 4 compares the proposed methods with both reference systems.

  • 13.6% test-set top 5 error was achieved by the final system, improving on the previous state of the art of 16.4%.
  • 10 neural networks composed the submitted system: 5 base models and 5 high-resolution models.
  • Table 4 compares the proposed methods with the previous year’s winner, Krizhevsky et al., and the current year’s winner, Clarifai.

6 Conclusion

The paper presents several improvements to neural-network image classification, spanning training augmentation, test-time prediction, and complementary high-resolution models.

  • Training-set augmentation used broader image crops and additional color manipulations to increase the effective training-set size.
  • Test-time augmentation generated predictions across multiple image scales and views, with a greedy algorithm reducing the number of predictions required.
  • Higher-resolution models generated useful complementary predictions through an efficient training approach.
  • A single base model and a single high-resolution model were as good as five base models.
  • The improvements were described as easy to implement and potentially applicable to other convolutional neural network image-classification systems.
Loading 1312.5402v1…