# A lightweight model and corn leaf disease recognition

## TL;DR

This paper introduces a lightweight model for accurately identifying corn leaf diseases, improving efficiency for use on portable devices.

## Contribution

The novel ES-ShuffleNetV2 model combines improved attention mechanisms and activation functions for efficient and accurate corn leaf disease recognition.

## Key findings

- The ES-ShuffleNetV2 model achieved 97.07% recognition accuracy, outperforming the base model.
- The model reduced parameters and FLOPs by 30.45% and 30.26%, respectively, after pruning.
- It outperformed competing models in Accuracy and F1-Score metrics.

## Abstract

Corn is a critical food crop globally, widely cultivated due to its strong adaptability. However, it is susceptible to various diseases, necessitating advanced intelligent detection methods to enhance disease prevention, control efficacy, and production efficiency. Traditional disease recognition models suffer from high computational costs or inadequate feature extraction capabilities, making it challenging to achieve efficient and accurate disease identification in complex environments. To improve the accuracy and efficiency of corn leaf disease identification and to meet the requirements of portable devices, this paper proposes a novel ES-ShuffleNetV2 (Exponential Linear Unit + Spatial Group-wise Squeeze-and-Excitation Block) lightweight recognition model for corn diseases. The proposed model builds upon the ShuffleNetV2 architecture. Firstly, an improved attention mechanism, SGSE, is incorporated immediately following the first convolutional layer to emphasize fine-grained features in corn leaf disease images, enhancing the model’s focus on key characteristics. Secondly, the model replaces the ReLU activation function in the down-sampling and basic units with the ELU function, facilitating smoother gradient propagation and faster convergence by allowing a small negative gradient inflow. Additionally, layer pruning techniques are employed to eliminate redundant parameters, reduce model complexity, and enhance operational efficiency on mobile devices. Experimental results demonstrated that the ES-ShuffleNetV2 model achieved recognition accuracy of 97.07%, surpassing the base model’s accuracy of 95.43%. After pruning, the new model reduced parameters by 30.45% and FLOPs by 30.26% compared to the original model, meeting the criteria for a lightweight recognition model. Furthermore, the ES-ShuffleNetV2 model outperformed competing models in Accuracy and F1-Score, validating its effectiveness in corn leaf disease recognition and providing valuable insights for future research.