A novel multiplier-less convolution core for YOLO CNN ASIC implementation

The You Only Look Once (YOLO) algorithm has a good trade-off between accuracy and execution speed in object detection. The main bottleneck of execution speed in YOLO is the optimum implementation of the Convolutional Neural Network (CNN). Reducing convolution core resources to increase parallelism c...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Journal of real-time image processing Jg. 21; H. 2; S. 45
Hauptverfasser: Bagherzadeh, Shoorangiz, Daryanavard, Hassan, Semati, Mohammad Reza
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Berlin/Heidelberg Springer Berlin Heidelberg 01.04.2024
Springer Nature B.V
Schlagworte:
Accuracy
Algorithms
Artificial neural networks
Circuit design
Computer Graphics
Computer Science
Image Processing and Computer Vision
Multimedia Information Systems
Multipliers
Neural networks
Object recognition
Pattern Recognition
Power consumption
Signal,Image and Speech Processing
Software
Multiplier-less
YOLO object detection
PE ASIC implementation
Convolutional neural network
ISSN:1861-8200, 1861-8219
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The You Only Look Once (YOLO) algorithm has a good trade-off between accuracy and execution speed in object detection. The main bottleneck of execution speed in YOLO is the optimum implementation of the Convolutional Neural Network (CNN). Reducing convolution core resources to increase parallelism can significantly increase the execution speed of the Algorithm. A new ASIC Processing Element (PE) is presented in this paper to reduce power consumption and increase speed while utilizing fewer resources. A multiplier-less convolution core is proposed by replacing multipliers with multiplexer circuits and designing a 19-input adder. Reducing the weight word length to five bits and compensating for the accuracy with the new quantization, has made the accuracy of the new architecture competitive with previous works. Compared with the traditional convolutional core, the best-proposed core has been improved by 4.44X, 4.9X, and 32% in power consumption, area, and delay, respectively. Placing the proposed core in the PE, the power consumption, FPS, and accuracy were 1.76W, 55.8, and 78%, respectively. Although the proposed 3 × 3 convolution core was evaluated using YOLOv2 and YOLOv4-tiny, it is also applicable to YOLOv7 and YOLOv8.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1861-8200
1861-8219
DOI:10.1007/s11554-024-01419-7