An Approach Using Threshold-Based Noise Reduction and Fine-Tuned ShuffleNetV2 for Plant Leaf Disease Detection

Diagnosing plant leaf diseases is essential for agricultural development. Leaves are an important part of the plant and are often where signs of disease appear. With the support of image processing algorithms, researchers have widely used them to support disease detection tasks on plant leaves. Transfer learning approaches have revealed encouraging results in many domains but require fine-tuning hyperparameter values. Additionally, a combination with noise reduction can lead to positive potential effects in improving performance. This study proposes an approach leveraging a noise reduction technique based on Soft-Thresholding with Lasso regression and then performing the disease classification with a fine-tuned ShuffleNetV2. The experimental results on 14,400 images of 24 plant leaf disease classes of 10 various plant species show that the Threshold-based noise reduction combined with a fine-tuned ShuffleNetV2 can obtain better performance in disease classification on plant leaves than the original model and several considered transfer learning methods.