Lightweight RepVGG-Based Cross-Modality Data Prediction Method for Solid Rocket Motors

Yang, Huixin; Zheng, Shangshang; Wang, Xu; Xu, Mingze; Li, Xiang

doi:10.3390/s23229165

Cited by 1 publication

(1 citation statement)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The model achieves an accuracy of over 80% on the ImageNet dataset and runs more than 53% faster compared to other models. Building upon this foundation, numerous scholars have extended its application to diverse fields such as forestry, agriculture, industry, and aerospace, yielding promising results [33][34][35][36].…”

Section: Introductionmentioning

confidence: 99%

A-RepVGG: Research on Classification Algorithms based on Deep Learning and Wood CT Images

Zheng,

Ge,

Yang

et al. 2024

Preprint

Self Cite

View full text Add to dashboard Cite

To address the issue of limited expressive ability and performance degradation of the model caused by the limited depth and width of the CNN because of low computational overhead of lightweight convolutional neural networks, this paper introduces a classification method called A-RepVGG for wood CT images. A-RepVGG aims to enhance the model's classification accuracy in terms of wood microstructure by increasing the model complexity without increasing the depth and width of the network. The method utilizes adaptive convolution to dynamically aggregate convolution kernels based on the input image. This allows the convolution kernels to have optimal receptive fields on different layers of features, enabling a more comprehensive extraction of information, such as wood texture, tubular pore distribution, and cell arrangement. Additionally, a multi-scale null attention mechanism is incorporated into the network, multi-layer convolution is employed to extract feature maps of different scales and weighted fusion is adopted to emphasize important feature regions. This effectively captures both local and global information of the image. Finally, the ELU activation function is introduced to ensure that the feature information can be properly output in the negative half-axis, thereby facilitating a more thorough extraction of wood feature information and improving the model's classification accuracy. The study aimed to classify 20 species of wood cross-section microscopic images. The findings revealed that the A-RepVGG model outperformed other existing wood image classification models, such as ResNet, ResNeSt, and ViT. The A-RepVGG model achieved an impressive accuracy of 99.50% on the test set and 99.20% on the validation set. This model incorporates adaptive convolution and attention mechanisms, which effectively enhance the classification accuracy of wood microscopic images. These results highlight the potential of deep learning in automatic classification of wood species and provide valuable insights for micro-level wood classification.

show abstract

Section: Introductionmentioning

confidence: 99%