Treffer: Variable convolution kernel with feature fusion and transfer learning for leaf disease classification.

The advancement of automated frameworks for detecting and classifying leaf diseases is extensively explored in contemporary agricultural practices. The effectiveness of a classifier relies on the feature extraction process. A novel Variable Convolution Kernel (VCK) feature extraction algorithm with Feature Fusion (FF) and Transfer Learning (TL) based disease classification is proposed. Sparse representation is obtained in the training stage by fusing features obtained through different filters. TL offers the benefit of leveraging pre-trained models on large datasets, saving significant time and computational resources when building and training new models for specific tasks. Mobilenet_v2 pretrained using ImageNet dataset can improve model performance, especially when dealing with limited training data, by transferring weights and features. A novel framework has been developed by incorporating CNN, TL, FF and tuning the hyperparameters. The underlying algorithm is known as FF-TL-CNN algorithm. The empirical investigation utilized the Plant Village dataset. The leaf disease categories examined in this study encompass early blight, black rot, bacterial spot, apple scab, cercospora leaf spot, and the category of healthy leaves. FF-TL-CNN outperformed other classifiers by attaining 98.85% accuracy, 98.63% precision, 98.41% recall and 99.32% F1-score with Plant Village dataset. The research findings demonstrate that the suggested deep learning model and algorithm have practical applications in real-world computer vision contexts, particularly in the field of agriculture. [ABSTRACT FROM AUTHOR]