Quality Grading of Oudemansiella raphanipes Using Three-Teacher Knowledge Distillation with Cascaded Structure for LightWeight Neural Networks

Agriculture, Journal Year: 2025, Volume and Issue: 15(3), P. 301 - 301

Published: Jan. 30, 2025

Oudemansiella raphanipes is valued for its rich nutritional content and medicinal properties, but traditional manual grading methods are time-consuming labor-intensive. To address this, deep learning techniques employed to automate the process, knowledge distillation (KD) used enhance accuracy of a small-parameter model while maintaining low resource occupation fast response speed in resource-limited devices. This study employs three-teacher KD framework investigates three cascaded structures: parallel model, standard series with residual connections (residual-series model). The student lightweight ShuffleNet V2 0.5x, teacher models VGG16, ResNet50, Xception. Our experiments show that structures result improved performance indices, compared ensemble equal weights; particular, residual-series outperforms other models, achieving 99.7% on testing dataset an average inference time 5.51 ms. findings this have potential broader application environments automated quality grading.

Language: Английский

---

Application of a Multi-Teacher Distillation Regression Model Based on Clustering Integration and Adaptive Weighting in Dam Deformation Prediction

Water, Journal Year: 2025, Volume and Issue: 17(7), P. 988 - 988

Published: March 27, 2025

Deformation is a key physical quantity that reflects the safety status of dams. Dam deformation influenced by multiple factors and has seasonal periodic patterns. Due to challenges in accurately predicting dam with traditional linear models, deep learning methods have been increasingly applied recent years. In response problems such as an excessively long training time, too-high model complexity, limited generalization ability large number complex hybrid models current research field, we propose improved multi-teacher distillation network for regression tasks improve performance model. The constructed using Transformer considers global dependencies, while student Temporal Convolutional Network (TCN). To efficiency, draw on concept clustering integration reduce teacher networks loss function tasks. We incorporate adaptive weight module into assign more teachers accurate prediction results. Finally, knowledge information formed based differences between network. concrete-faced rockfill located Guizhou province, China, results demonstrate that, compared other methods, this approach exhibits higher accuracy practicality.

Language: Английский

---

Class-adaptive attention transfer and multilevel entropy decoupled knowledge distillation

Multimedia Systems, Journal Year: 2025, Volume and Issue: 31(3)

Published: April 15, 2025

Language: Английский

---

A Study on the Change of Artistic Creation Styles in the Internet Era in Higher Art Education Based on the Perspective of Network Analysis

Applied Mathematics and Nonlinear Sciences, Journal Year: 2024, Volume and Issue: 9(1)

Published: Jan. 1, 2024

Abstract Artistic creation in the network era, examined terms of artistic style, can be roughly divided into several important stages time, such as 1985-1995, 1996-2005, 2006-2015, and 2016-present. This paper integrates pedagogical knowledge distillation an style feature extraction model to investigate change era. Using global features, color histogram art image region is extracted by calculating features on H, S, V from both texture features. The vectors are normalized, assuming that they conform a Gaussian distribution. Quantify using adaptive weighted Gram matrix improve accuracy classification. Joint for dissemination Plotting scatter chromaticity values L*-b* all painting samples reveals luminance L* paintings fluctuates between 21.38 79.94, red/greenness value a* -8.57 1.7, yellow/blueness b* 1.67 16.21. temporal development characteristics styles follows clear pattern. In comparison eigenvalues under variable parameters, moment inertia images period 1985-1995 raised 147.698 615.965 after noise addition, there different degrees growth other periods, but with magnitudes growth, obvious differences.

Language: Английский

---