Optimizing FCN for devices with limited resources using quantization and sparsity enhancement

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Resource-limited devices pose challenges for deploying complex deep learning models like fully convolutional networks (FCNs).
• Prior research focused on quantizing architectures like VGG-16, with limited exploration of comprehensive layer-wise quantization in FCN-8.

Purpose of the Study:

• To optimize FCNs for real-time deployment on resource-constrained devices.
• To investigate comprehensive layer-wise quantization techniques for the FCN-8 architecture.

Main Methods:

• Proposed an innovative approach using full-layer quantization with an error minimization algorithm.
• Employed sensitivity analysis to optimize fixed-point representation of network weights.
• Utilized retraining to maintain network performance post-quantization.

Main Results:

• Achieved significant network sparsity, up to 40%, under extreme quantization conditions.
• Preserved network performance, yielding 89.3% pixel accuracy.
• Demonstrated effectiveness in both image classification and semantic segmentation tasks.

Conclusions:

• Full-layer quantization and retraining are effective for reducing network complexity.
• This approach successfully maintains high accuracy in FCNs for real-time applications on resource-limited devices.