Efficient Integer Quantization for Compressed DETR Models.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• DETR (DEtection TRansformer) offers powerful object detection but has high computational and storage demands.
• Resource-constrained devices face deployment challenges due to DETR's extensive requirements.

Purpose of the Study:

• To develop an efficient and hardware-friendly target detection model suitable for resource-constrained environments.
• To reduce the computational and storage complexity of the DETR model.

Main Methods:

• Replaced the ResNet-50 backbone in DETR with Swin-T for unified Transformer processing.
• Proposed a novel integer-only quantization scheme for all non-linear layers (Sigmoid, Softmax, LayerNorm, GELU) in addition to linear layers.
• Implemented a data compression method to decrease memory occupation and computational complexity.

Main Results:

• Reduced model computation to 6.3% and storage to 25% of the original DETR model.
• Achieved a minimal average accuracy decrease of only 1.1%.
• Demonstrated the effectiveness of integer-only quantization across all operational layers.

Conclusions:

• The proposed method significantly enhances the efficiency of DETR for deployment on resource-constrained devices.
• Integer-only quantization of all layers offers a viable approach to reduce model size and computation without substantial accuracy degradation.
• This approach presents a practical solution for efficient hardware-based target detection.