[Paper Review] Some Improvements on Deep Convolutional Neural Network Based Image Classification

Andrew Howard|arXiv (Cornell University)|Dec 19, 2013

Advanced Neural Network Applications7 references225 citations

TL;DR

This paper introduces improvements to deep convolutional neural networks for image classification, focusing on data augmentation through additional training transformations, test-time ensemble predictions via data augmentation, and multi-scale inference using higher-resolution inputs. The proposed methods reduced the top-5 error rate to 13.55% on ImageNet without external data, achieving over a 20% relative improvement over the previous year's winner.

ABSTRACT

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.

Motivation & Objective

• To improve deep convolutional neural network performance on image classification tasks using internal data and model enhancements.
• To reduce overfitting and improve generalization through extensive data augmentation during training.
• To enhance prediction robustness by generating multiple predictions at test time using augmented inputs.
• To leverage higher-resolution image processing to improve feature representation and classification accuracy.
• To achieve state-of-the-art performance in the ImageNet Large Scale Visual Recognition Challenge 2013 using only internal data.

Proposed method

• Applied additional image transformations during training, such as random crops, color distortions, and horizontal flips, to increase data diversity.
• Used data augmentation at test time to generate multiple predictions from the same input image, then averaged or ensembled the results.
• Employed complementary models trained on different image scales, with higher-resolution inputs to capture finer-grained features.
• Combined multiple models with diverse architectures and training strategies to improve overall robustness and accuracy.
• Optimized training with stochastic gradient descent and learning rate decay, using batch normalization and ReLU activations.
• Used a multi-scale inference strategy where the same image was classified at multiple resolutions, and results were fused to improve final predictions.

Experimental results

Research questions

• RQ1 How can data augmentation during training improve generalization in deep convolutional neural networks for image classification?
• RQ2 To what extent does test-time data augmentation improve prediction accuracy and robustness?
• RQ3 Can combining predictions from models applied to different image resolutions enhance classification performance?
• RQ4 How effective are complementary models in reducing top-5 error rates without using external data?
• RQ5 What is the maximum performance improvement achievable using only internal data and architectural enhancements?

Key findings

• The system achieved a top-5 error rate of 13.55% on the ImageNet validation set using only internal training data.
• The improvement over the previous year’s winner was over 20% relative reduction in top-5 error rate.
• Test-time data augmentation significantly improved prediction robustness and accuracy by generating multiple predictions per input.
• Multi-scale inference using higher-resolution images contributed to better feature learning and improved classification outcomes.
• The combination of data augmentation, test-time ensembling, and multi-scale processing led to state-of-the-art performance in the 2013 ImageNet challenge.
• The method demonstrated strong generalization without relying on external datasets, highlighting the effectiveness of internal data augmentation and model ensembling.