Intermediate Deep Feature Compression: Toward Intelligent Sensing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Advances in hardware technology enable practical deep learning at the front-end.
• Current methods involve transmitting raw visual data or high-level features, which can be inefficient.
• Intelligent sensing at the front-end requires optimized data representation for cloud-based analysis.

Purpose of the Study:

• To propose a novel strategy for compactly representing and conveying intermediate-layer deep learning features.
• To facilitate a collaborative approach between front-end intelligent sensing and cloud-based analysis.
• To achieve a balance between computational load, transmission load, and generalization ability for deep neural networks.

Main Methods:

• Representing and transmitting intermediate-layer deep learning features instead of raw data or top-layer features.
• Evaluating both lossless and lossy compression techniques for these intermediate features.
• Investigating the feasibility of standardizing deep feature coding for multi-task benefits.

Main Results:

• The proposed strategy effectively balances computational and transmission loads for cloud-based visual analysis.
• Intermediate-layer feature representation demonstrates high generalization capability.
• Evaluations provide baselines for lossless and lossy deep feature compression.

Conclusions:

• Compact representation of intermediate-layer deep learning features is a promising strategy for front-end and cloud collaboration.
• This approach supports large-scale cloud-based visual analysis with improved efficiency and generalization.
• The findings pave the way for standardization activities in deep feature coding.