Hyperparameter optimization ResNet by improved Beluga Whale Optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Neural network performance is highly dependent on parameter tuning.
• Existing ResNet models face challenges with accuracy and parameter optimization.
• Swarm intelligence algorithms offer potential for improving optimization processes.

Purpose of the Study:

• To develop an improved Beluga Whale Optimization (EBWO) algorithm by incorporating the Tent chaotic map.
• To construct a novel EBWO-ResNet model by integrating EBWO with the ResNet architecture for enhanced performance.
• To evaluate the efficacy of the EBWO algorithm and the EBWO-ResNet model in engineering problems and maize disease identification.

Main Methods:

• The Tent chaotic map was introduced into the Beluga Whale Optimization algorithm to create the EBWO algorithm, addressing initial population limitations.
• The EBWO algorithm was integrated with the ResNet model to form the EBWO-ResNet model, targeting improved accuracy and parameter tuning.
• The EBWO algorithm was tested on three engineering problems against five other algorithms. The EBWO-ResNet model was applied to maize disease identification and compared with seven other models.

Main Results:

• The EBWO algorithm demonstrated superior performance in solving the three engineering problems compared to five other swarm intelligent algorithms.
• The EBWO-ResNet model achieved a high accuracy of 96.3% in maize disease identification.
• The EBWO-ResNet model outperformed seven other comparative models in maize disease identification by 0.2-1.5 percentage points.

Conclusions:

• The proposed EBWO algorithm effectively enhances the optimization process.
• The EBWO-ResNet model significantly improves the accuracy of maize disease identification, contributing to better crop yield management.
• The developed EBWO-ResNet model shows strong potential for practical applications in agricultural disease detection.