An FPGA-Based Reconfigurable CNN Accelerator for YOLO

Zhang, Shiguang; Cao, Jian; Zhang, Quan; Zhang, Qi; Zhang, Ying; Wang, Yuan

doi:10.1109/icet49382.2020.9119500

Cited by 68 publications

(62 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Weight size (MB) Original YOLOv3 [10] 54.64 32 237 YOLOv2 [34] 48.1 32 194 Zhang et al [40] 48.1 16 97 YOLOv3-tiny [10] 33.1 32 33.79 Gaussian YOLOv3-tiny [13] 39.…”

Section: Methods Map (%) W Bitsmentioning

confidence: 99%

“…Table 6 compares the proposed network with other networks in terms of accuracy and weight size. Compared to YOLOv2 network [34] and its quantized model [40], the proposed network achieves 6.42 higher mAP even though the weight size of the proposed network is approximately onequarter and one-half smaller, respectively. Compared to YOLOv3-tiny [10], Gaussian YOLOv3-tiny [13], and Nayak et al [41], the weight size of the proposed network is approximately one-third larger, but the mAP of the proposed network is much higher than that of YOLOv3-tiny [10], Gaussian YOLOv3-tiny [13], and Nayak et al [41] by 21.42, 15.22, and 26.43, respectively.…”

Section: Performance Evaluation On Object Detectionmentioning

confidence: 96%

See 1 more Smart Citation

Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors

Kim

2021

IEEE Access

View full text Add to dashboard Cite

In the field of object detection, deep learning has greatly improved accuracy compared to previous algorithms and has been used widely in recent years. However, object detection using deep learning requires many hardware (HW) resources due to the huge computations for high performance, making it very difficult to run real-time on embedded platforms. Therefore, various compression methods have been studied to solve this problem. In particular, quantization methods greatly reduce the computational burden of deep learning by reducing the number of bits used for weights and activation functions in deep learning. However, most of these existing studies targeted only object classification and cannot be applied to object detection. Furthermore, most of the existing quantization studies are based on floating-point operations, which requires additional effort when implementing HW accelerators. This paper proposes an HW-friendly fixed-point-based quantization method that can also be applied to object detection. In the proposed method, the center of the weight distribution is adjusted to zero by subtracting the mean of weight parameters before quantization, and the retraining process is iteratively applied to minimize the accuracy drop caused by quantization. Furthermore, while applying the proposed method to object detection, performance degradation is minimized by considering the minimum and maximum values of weight parameters of deep learning networks. When applying the proposed quantization method to representative one-stage object detectors, You Only Look Once v3 and v4 (YOLOv3 and YOLOv4), detection accuracy similar to the original networks (i.e., YOLOv3 and YOLOv4) with a single-precision floating-point format (32-bit) is maintained despite expressing weights with only about 20% of the bits compared to a singleprecision floating-point format in COCO dataset. INDEX TERMS Convolutional neural network, deep neural network, fixed-point quantization, network compression, object detector, YOLOv3, YOLOv4.

show abstract

“…Weight size (MB) Original YOLOv3 [10] 54.64 32 237 YOLOv2 [34] 48.1 32 194 Zhang et al [40] 48.1 16 97 YOLOv3-tiny [10] 33.1 32 33.79 Gaussian YOLOv3-tiny [13] 39.…”

Section: Methods Map (%) W Bitsmentioning

confidence: 99%

Section: Performance Evaluation On Object Detectionmentioning

confidence: 96%

Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors

Kim

2021

IEEE Access

View full text Add to dashboard Cite

show abstract

“…Conventionally, various AI inference schemes for object detection have been proposed, such as you-onlylook-once (YOLO) [3][4][5] and single-shot-multiboxdetection (SSD) [6], and dedicated engines have also been proposed for executing these schemes in real time [7][8][9][10][11][12][13][14][15][16][17][18][19][20]. The inference scheme has a convolutional neural network (CNN) that predicts object bounding boxes and class probabilities in a single evaluation.…”

Section: Introductionmentioning

confidence: 99%

High-definition object detection technology based on AI inference scheme and its implementation

Uzawa

Yoshida

Iinuma

et al. 2021

IEICE Electron. Express

View full text Add to dashboard Cite

Video artificial intelligence (AI) applications for edges/terminals, such as non-visual drone flight, require object detection in high-definition video in order to detect a wide range of objects with one camera at once. To enable satisfying this requirement, we propose a new high-definition object detection technology based on an AI inference scheme and its implementation. In the technology, multiple object detectors cooperate to detect small and large objects in high-definition video. The evaluation results show that our technology can achieve 2.1 times higher detection performance in full HD images thanks to the cooperation of three object detectors.

show abstract

“…Zhang el al. [32] implemented the YOLOv2 and YOLOv2-tiny models in an FPGA and evaluated its performance.…”

Section: Introductionmentioning

confidence: 99%

Design of a Scalable and Fast YOLO for Edge-Computing Devices

Han

Lee

Lim

et al. 2020

Sensors

View full text Add to dashboard Cite

With the increase in research cases of the application of a convolutional neural network (CNN)-based object detection technology, studies on the light-weight CNN models that can be performed in real time on the edge-computing devices are also increasing. This paper proposed scalable convolutional blocks that can be easily designed CNN networks of You Only Look Once (YOLO) detector which have the balanced processing speed and accuracy of the target edge-computing devices considering different performances by exchanging the proposed blocks simply. The maximum number of kernels of the convolutional layer was determined through simple but intuitive speed comparison tests for three edge-computing devices to be considered. The scalable convolutional blocks were designed in consideration of the limited maximum number of kernels to detect objects in real time on these edge-computing devices. Three scalable and fast YOLO detectors (SF-YOLO) which designed using the proposed scalable convolutional blocks compared the processing speed and accuracy with several conventional light-weight YOLO detectors on the edge-computing devices. When compared with YOLOv3-tiny, SF-YOLO was seen to be 2 times faster than the previous processing speed but with the same accuracy as YOLOv3-tiny, and also, a 48% improved processing speed than the YOLOv3-tiny-PRN which is the processing speed improvement model. Also, even in the large SF-YOLO model that focuses on the accuracy performance, it achieved a 10% faster processing speed with better accuracy of 40.4% mAP@0.5 in the MS COCO dataset than YOLOv4-tiny model.

show abstract

An FPGA-Based Reconfigurable CNN Accelerator for YOLO

Cited by 68 publications

References 8 publications

Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors

Zero-Centered Fixed-Point Quantization With Iterative Retraining for Deep Convolutional Neural Network-Based Object Detectors

High-definition object detection technology based on AI inference scheme and its implementation

Design of a Scalable and Fast YOLO for Edge-Computing Devices

Contact Info

Product

Resources

About