Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks

Chen, Jierun; Kao, Shiu-hong; He, Hao; Zhuo, Weipeng; Wen, Song; Lee, Chul-Ho; Chan, S.-H. Gary

doi:10.1109/cvpr52729.2023.01157

Cited by 693 publications

(125 citation statements)

References 31 publications

Supporting

Mentioning

125

Contrasting

Order By: Relevance

“…1 × 1 convolution, and this cheap operation inevitably affects the feature representation capability. Therefore, in this paper, a Dy-G module is constructed based on GhostNet by introducing dynamic convolution [25] to enhance feature representation and PConv [20] to balance the computational costs. It is constructed as shown in figure 7.…”

Section: Construction Of Dy-g Modulementioning

confidence: 99%

“…MobileNets [13][14][15], ShuffleNets [16,17], introduce depth-separable convolution and channel shuffling operations respectively to further reduce the number of floating-point operations (FLOPs). GhostNet [18,19] and FasterNet [20] both propose new convolution operators from the perspective of feature graph redundancy to build lightweight network architectures. In the field of fault diagnosis, there are also many lightweight designs with the help of the above ideas.…”

Section: Introductionmentioning

confidence: 99%

“…At the same time, inspired by the design of Ghost convolution and taking the redundancy between feature maps as a starting point, we design Dy-GNet, which has been accomplished to model not only a more streamlined structure, but also a fast response time and higher accuracy. Specifically the Dy-G module is designed to improve the feature extraction capability without increasing the FLOPs by replacing the conventional convolution with dynamic convolution [25], and the computational speed of the convolution operation is improved by replacing the depth-wise convolution (DWConv) with partial convolution (PConv) [20], which convolves the features of only some channels. Meanwhile, compared with other common models, we introduce the wavelet scattering transform [26] (WST) to approximately reduce the original data, which not only extracts the time-frequency information of the data, but also reduces the data input and improves the efficiency of model recognition.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Dy-GNet: a lightweight and efficient 1DCNN-based network for leakage aperture identification

Zhang,

Li,

Liu

et al. 2024

Meas. Sci. Technol.

View full text Add to dashboard Cite

The deep neural network's gas leakage aperture recognition method has high accuracy, but its low computational efficiency due to its complex structure greatly limit its application in resource-limited industrial environments and real-time processing. In this paper, we proposed an efficient recognition method based on 1D convolutional neural network. First, wavelet scattering coefficients with time-frequency information are obtained using the wavelet scattering transform. Second, the dynamic convolution is used to deepen the feature extraction and PConv to speed up the inference time, and the efficient Dy-G module is constructed. Finally, the Dy-G module is stacked to construct Dy-GNet to achieve the classification task. The performance of the model is verified in a noisy environment, and the results show that the model FLOPs are 17.24 M and 83.94% accuracy is achieved at SNR=0 dB, which guarantees the accuracy while speeding up the inference speed with high efficiency.

show abstract

Section: Construction Of Dy-g Modulementioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Dy-GNet: a lightweight and efficient 1DCNN-based network for leakage aperture identification

Zhang,

Li,

Liu

et al. 2024

Meas. Sci. Technol.

View full text Add to dashboard Cite

show abstract

“…This is mostly because of the cucumber pruning targets' severe occlusion, the wide disparity in target size, and the complicated background. We construct Faster Cross-Stage Partial Network (FCSP) by combining Faster Net [9] , Batch Normalization (BN, and the C3 module to address the aforementioned issues. The structure of FCSP is shown in Fig.…”

Section: Structure Of Faster Cross Stage Partial Network (Fcsp)mentioning

confidence: 99%

LE-YOLO: a lightweight and efficient cucumber pruning target detection and recognition method based on improved YOLOv5s

Huang,

2023

Fourth International Conference on Signal Processing and Computer Science (SPCS 2023)

View full text Add to dashboard Cite

“…In order to speed up the feature extraction and processing efficiency of the backbone network for input images, this paper introduces the newly proposed structure PConv (Partial Convolution) in CVPR 2023 [8]. PConv is proposed to solve the computational inefficiency of the neural network, and by reducing both redundant computation and memory access, FLOPS can be more high.…”

Section: The Pconv Modulementioning

confidence: 99%

A laser welding defect detection model based on improved YOLOv5

Yang,

Zhou,

et al. 2023

Sixth International Conference on Computer Information Science and Application Technology (CISAT 2023)

View full text Add to dashboard Cite

With the expanding market demand for new energy power batteries, the industrial requirement for detecting laser welding defects of lithium battery’s pole are also increasing. Since the traditional machine vision defect detection has the problems of low detection accuracy and slow speed, this paper proposes a laser welding defect detection model based on the improved YOLOv5 algorithm to meet the requirements of industrial production. First, all C3 modules of the backbone was substituted for the newly proposed PConv module, which not only reduces the amount of redundant computation, improves the detection speed, but also increases the feature extraction efficiency of the backbone network for input images; second, we add the attention mechanism SE to the penultimate layer of the backbone network, which enhances the useful features of the backbone network for input images without additionally increasing the computation. This approach enhances the backbone network's ability to extract useful features from the input images and improves the accuracy of the model. The experimental results show that the improved model improves the average detection accuracy by 1.1% to 97% compared with the original YOLOv5 model. The FPS reaches 333, which can meet the current industrial demand for real-time defect detection.

show abstract

Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks

Cited by 693 publications

References 31 publications

Dy-GNet: a lightweight and efficient 1DCNN-based network for leakage aperture identification

Dy-GNet: a lightweight and efficient 1DCNN-based network for leakage aperture identification

LE-YOLO: a lightweight and efficient cucumber pruning target detection and recognition method based on improved YOLOv5s

A laser welding defect detection model based on improved YOLOv5

Contact Info

Product

Resources

About