Compression Helps Deep Learning in Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) models for image classification typically use pristine images.
• JPEG compression is conventionally understood to degrade DL classification accuracy.
• The impact of selecting specific JPEG quality factors on DL performance is not well-established.

Purpose of the Study:

• To investigate the effect of JPEG compression on deep neural network (DNN) image classification accuracy.
• To demonstrate that strategic selection of JPEG compressed images can improve classification accuracy and reduce data size.
• To develop methods for selecting optimal JPEG versions for input into DNNs.

Main Methods:

• Utilized 10 JPEG compressed versions (Quality Factors 10-100) of original images.
• Introduced the Highest Rank Selector (HRS) for selecting optimal JPEG inputs when ground truth is known.
• Proposed a novel convolutional neural network (CNN) topology with 11 parallel inputs for cases where ground truth is unknown.

Main Results:

• HRS improved Top-1 accuracy by 5.6% and Top-5 by 1.9% on ImageNet, with a compression ratio (CR) of 8.
• The novel CNN topology improved Top-1 accuracy by ~0.4% and Top-5 by ~0.32% (Inception V3) and ~0.2% (ResNet-50 V2).
• Alternative selectors maintained accuracy while achieving CRs up to 8.8.

Conclusions:

• JPEG compression, when strategically applied, can enhance DL image classification performance.
• The proposed HRS and CNN methods offer effective ways to leverage JPEG compression for improved accuracy and efficiency.
• This research challenges the notion that JPEG compression universally harms DL classification accuracy.