A Reconfigurable Coarse-to-Fine Approach for the Execution of CNN Inference Models in Low-Power Edge Devices
Convolutional neural networks (CNNs) have evolved into essential components for a wide range of embedded applications due to their outstanding efficiency and performance. To efficiently deploy CNN inference models on resource-constrained edge devices, field programmable gate arrays (FPGAs) have beco...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wiley
2024-01-01
|
| Series: | IET Computers & Digital Techniques |
| Online Access: | http://dx.doi.org/10.1049/cdt2/6214436 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Convolutional neural networks (CNNs) have evolved into essential components for a wide range of embedded applications due to their outstanding efficiency and performance. To efficiently deploy CNN inference models on resource-constrained edge devices, field programmable gate arrays (FPGAs) have become a viable processing solution because of their unique hardware characteristics, enabling flexibility, parallel computation and low-power consumption. In this regard, this work proposes an FPGA-based dynamic reconfigurable coarse-to-fine (C2F) inference of CNN models, aiming to increase power efficiency and flexibility. The proposed C2F approach first coarsely classifies related input images into superclasses and then selects the appropriate fine model(s) to recognise and classify the input images according to their bespoke categories. Furthermore, the proposed architecture can be reprogrammed to the original model using partial reconfiguration (PR) in case the typical classification is required. To efficiently utilise different fine models on low-cost FPGAs with area minimisation, ZyCAP-based PR is adopted. Results show that our approach significantly improves the classification process when object identification of only one coarse category of interest is needed. This approach can reduce energy consumption and inference time by up to 27.2% and 13.2%, respectively, which can greatly benefit resource-constrained applications. |
|---|---|
| ISSN: | 1751-861X |