Scholarlink

<p>The widespread use of Convolutional Neural Networks (CNNs) in practical applications has fueled a growing demand for light-weight CNNs deployable on resource-constrained devices. In this paper, an efficient light-weight CNN called CSPMobileNet is proposed based on depth-wise separable convolutions (DWConv), split-and-merge strategy, and inverted residual structure. The proposed network is composed of some CSPConv blocks to reduce the model size and computational operations. Each CSPConv block is designed based on the split-and-merge strategy and has some inverted residual blocks. In addition, a CSPConv block with different attention mechanisms is analyzed in the proposed scheme. For performance evaluation, four public image datasets, Cifar10, Cifar100, Bird100, and Bird325, are utilized and some existing light-weight networks such as MobileNetV3, ShuffleNet, and EfficientNet are selected for testing. Here we perform the subjective and objective evaluation for performance analysis. For the subjective evaluation, the gradient-weighted class activation mapping (Grad-CAM) is utilized to understand the effect of feature extraction in the proposed scheme. The experimental results show that the proposed light-weight network can effectively extract representative visual features from images. As for the objective evaluation, the inference time of the proposed scheme is shorter than those of MobileNetV3, EfficientNetV1-B0, and ShuffleNetV2 on PC. Compared with MobileNetV3, the model size and inference time of the proposed network are also smaller on a resource-constrained device, Samsung Galaxy Tab S6 Lite with Exynos 9611. Therefore, the experimental results show that the proposed light-weight network, CSPMobileNet, can not only have a small model size but also achieve fast inference for image classification compared with MobileNetV3.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An efficient light-weight convolutional neural network based on split-and-merge strategy and inverted residual structure for resource-constrained devices

  • 期刊论文

  • 发表日期：

    2025年11月30日

查看全文
  • Siou-Min Lin,
  • Kuan-Ting Lai,
  • Guo-Shiang Lin,
  • Chuan-Wang Chang,
  • Ku-Yaw Chang

摘要

The widespread use of Convolutional Neural Networks (CNNs) in practical applications has fueled a growing demand for light-weight CNNs deployable on resource-constrained devices. In this paper, an efficient light-weight CNN called CSPMobileNet is proposed based on depth-wise separable convolutions (DWConv), split-and-merge strategy, and inverted residual structure. The proposed network is composed of some CSPConv blocks to reduce the model size and computational operations. Each CSPConv block is designed based on the split-and-merge strategy and has some inverted residual blocks. In addition, a CSPConv block with different attention mechanisms is analyzed in the proposed scheme. For performance evaluation, four public image datasets, Cifar10, Cifar100, Bird100, and Bird325, are utilized and some existing light-weight networks such as MobileNetV3, ShuffleNet, and EfficientNet are selected for testing. Here we perform the subjective and objective evaluation for performance analysis. For the subjective evaluation, the gradient-weighted class activation mapping (Grad-CAM) is utilized to understand the effect of feature extraction in the proposed scheme. The experimental results show that the proposed light-weight network can effectively extract representative visual features from images. As for the objective evaluation, the inference time of the proposed scheme is shorter than those of MobileNetV3, EfficientNetV1-B0, and ShuffleNetV2 on PC. Compared with MobileNetV3, the model size and inference time of the proposed network are also smaller on a resource-constrained device, Samsung Galaxy Tab S6 Lite with Exynos 9611. Therefore, the experimental results show that the proposed light-weight network, CSPMobileNet, can not only have a small model size but also achieve fast inference for image classification compared with MobileNetV3.