Abstract: 【Objective】To identify the most suitable mainstream image preprocessing method for deploying ConvNeXt convolutional neural network model on portable handheld devices in tobacco leaf maturity identification.【Methods】Tobacco leaf images were collected by a portable handheld image acquisition device. The images were preprocessed by four preprocessing methods,namely Gaussian scaling,contrast enhancement,color enhancement,and crop-scaling,respectively. A ConvNeXt convolutional neural network was used to construct a model,and the accuracy,training time,and model size of the model for tobacco leaf maturity identification were recorded. By comparative analysis of the performance,training efficiency,and model size of different preprocessing methods,evaluate the potential application of ConvNext convolutional neural network model in identifying tobacco leaf maturity on portable devices.【Results】Among the four image preprocessing methods,Gaussian scaling the best overall performance at tobacco leaf maturity identification when combined with the ConvNeXt convolutional neural network model. The model trained with Gaussian scaled images achieved an accuracy of 97.68%,outperforming the models trained with contrast enhancement,color enhancement,and crop-scaling. It also demonstrated the shortest training time（8.927 minutes） and a compact model size（63.5 MB）,highlighting both efficiency and lightweight characteristics. Compared to other models such as YOLO and XGBoost,the ConvNeXt model with Gaussian scaling showed superior performance across all indexes,especially in terms of accuracy and training speed,making it well-suited for deployment on portable handheld devices.【Conclusion】As an image preprocessing method,Gaussian scaling can effectively enhance the accuracy and operational efficiency of the ConvNeXt convolutional neural network model for tobacco leaf maturity identification tasks. The model constructed by combining Gaussian scaling with ConvNeXt convolutional neural network has fast training speed and low resource consumption,making it suitable for use on portable handheld image acquisition devices.

Key words: ConvNeXt convolutional neural network model, Tobacco leaf maturity identification, Portable handheld image acquisition devices, Intelligent image identification, Image preprocessing methods