Source-linked AI summary

Convolutional Neural Networks Applied to House Numbers Digit Classification

Pierre Sermanet, Soumith Chintala, Yann LeCun

arXiv:1204.3968v1cs.CVcs.LGcs.NE

TL;DR

Recognizing digits in natural-scene house-number photographs remains difficult because of low contrast, blur, low resolution, and illumination variation. The paper evaluates fully supervised ConvNets augmented with Lp pooling and multi-stage features on SVHN, achieving 94.85% accuracy and improving the previous 90.6% state of the art by 4.25 points.

  • Problem

    Digit recognition in natural-scene photographs is challenging because of non-contrasting backgrounds, low resolution, blur, and illumination differences.

  • Method

    The paper augments a traditional fully supervised ConvNet with Lp pooling, multi-stage features, contrast normalization, and a two-stage feature extractor.

  • Results

    94.85% accuracy improves the previous 90.6% state of the art by 4.25 points on SVHN.

  • Takeaways & Limitations

    Lp pooling performs best at intermediate p values, while multi-stage features provide only a slight increase for house-number classification.

  • Takeaways & Limitations

    The reported approach uses fully supervised training, whereas the best previous methods use unsupervised learning, so the contribution of supervision is not isolated.

Abstract

from arXiv · show

We classify digits of real-world house numbers using convolutional neural networks (ConvNets). ConvNets are hierarchical feature learning neural networks whose structure is biologically inspired. Unlike many popular vision approaches that are hand-designed, ConvNets can automatically learn a unique set of features optimized for a given task. We augmented the traditional ConvNet architecture by learning multi-stage features and by using Lp pooling and establish a new state-of-the-art of 94.85% accuracy on the SVHN dataset (45.2% error improvement). Furthermore, we analyze the benefits of different pooling methods and multi-stage features in ConvNets. The source code and a tutorial are available at eblearn.sf.net.

1. Introduction

Recognizing digits in natural-scene photographs is difficult because of challenging image conditions, motivating the SVHN benchmark and learned ConvNet features. The paper reports 94.85% accuracy, improving on the previous 90.6% state of the art.

  • Natural-scene digit recognition is harder than handwritten or typed character recognition because images may have low contrast, low resolution, blur, and varied illumination.
  • SVHN provides 600,000 labeled, color 32x32 digit images extracted from street-level house-number photographs with natural backgrounds.
  • ConvNets learn features from pixels through the classifier, contrasting with earlier approaches based on hand-crafted features and template matching.
  • 94.85% accuracy improves on the previous 90.6% state of the art by 4.25 points using fully supervised ConvNets with modified pooling and multi-stage features.

2 Architecture

The architecture stacks convolution, pooling, and normalization stages, augmenting traditional ConvNets with Lp pooling and multi-stage features. Multi-stage features add lower-level information, but their gains are minimal for house numbers relative to more textured or multi-scale objects.

  • 2 Architecture: Each feature stage contains convolution, pooling/subsampling, and normalization, with Lp pooling and subtractive-only normalization used instead of traditional pooling and divisive normalization.
  • 2.1 Lp-Pooling: L2 pooling applies a 3x3 Gaussian kernel with a 2x2 stride to a 9x9 feature map.
  • 2.2 Multi-Stage Features: Multi-stage features branch outputs from all stages into the classifier, combining complementary local textures and fine details with higher-level features.
  • 2.2 Multi-Stage Features: The first-stage features are branched, subsampled again, and concatenated with second-stage features before entering the classifier.
  • 2.2 Multi-Stage Features: Multi-stage features provide only minimal gains for house numbers compared with their improvements on pedestrians and traffic signs.

3. Experiments

The experiments define a validation split from SVHN’s train and extra sets, preprocess images with contrast normalization, and evaluate a two-stage ConvNet. Validation comparisons favor intermediate Lp-pooling values, while multi-stage improvements remain slight.

  • 3.1. Data Preparation: The 6,000-sample validation set combines 4,000 training samples and 2,000 extra samples, emphasizing difficult samples while retaining easy examples.
  • 3.1. Data Preparation: Images receive local contrast normalization on the Y channel followed by global contrast normalization across channels, without sample distortions.
  • 3.2 Architecture Details: The model uses two feature-extraction stages and a two-layer nonlinear classifier, with first-layer inputs added to reinforce global features with local motifs.
  • 3.2 Architecture Details: Multi-stage features yield only a slight validation error improvement over single-stage features for house-number classification.
  • 3.2 Architecture Details: Stochastic gradient descent is used for optimization, with the dataset shuffled after each training iteration.
  • 3.2 Architecture Details: p = 12 achieves a 5.61% validation error after 1000 training epochs, outperforming the other evaluated Lp-pooling values.

4 Results & Future Work

The experiments report state-of-the-art test accuracy with Lp pooling, while multi-stage features provide only a slight additional improvement. The authors identify fully supervised training and large scale variations as important boundaries for interpreting the results and guiding future work.

  • Results: 94.85% accuracy establishes state-of-the-art performance with L4 pooling, compared with 90.6% previously.
  • Results: Lp pooling with 1 < p < ∞ has a clear validation and test advantage over average pooling on this dataset.Average pooling is 3.58 points inferior to L2 pooling in Table 2.
  • Results: Multi-stage features produce only a slight performance increase relative to gains reported in other vision applications.
  • Future Work: The approach is trained fully supervised, whereas the best previous methods use unsupervised learning.The authors propose future experiments to determine how much accuracy improvement is attributable to supervision.
  • Future Work: Highest-error validation samples appear to exhibit large scale variations, motivating artificial scale deformations during training as future work.Figure 6 shows these samples for the 94.33% accuracy L2-pool based multi-stage ConvNet.
Loading 1204.3968v1…