An empirical evaluation of image preprocessing techniques for transfer learning-based multicrop classification
Аннотация
The accurate and correct classification of leaves as healthy is quite necessary in advanced precision agriculture. However, more work on the effects of preprocessing in machine learning is needed. The study worked on evaluating the various preprocessing techniques on five classes of healthy leaves, the dataset consisting of apple, blueberry, grape, cherry, and corn leaves. The dataset was split into train, test, and validation parts, then the train images were augmented with horizontal and vertical flips. Later preprocessed with five various approaches, i.e., raw data as a baseline, gamma enhancement, Sobel edge detection, histogram equalisation, and lastly, a brightness-adjusted preprocessing technique. Trained on various deep learning and transfer learning techniques, i.e., CNN-based architecture, AlexNet, DenseNet, InceptionV3, MobileNetV2, ResNet-50, and VGG-16, maintaining consistent splits and maintaining uniformity. Here, a total of 35 hybrid architectures are explored. Performance is being assessed using accuracy, precision, recall, MCC, F1 score, and AUC-ROC. AlexNet performed the best of the bunch with Raw images, sitting at 99% test accuracy, and MobileNet with a 99% accuracy on the Gamma-enhanced dataset, with a significant drop in Sobel edge detection-based datasets for certain models like Inception-V3 and the custom model. The findings of this study highlight how various preprocessing techniques affect training and the selection of appropriate preprocessing techniques to optimize deep learning models for excellent agricultural image classification, giving strong insights for automated crop species detection systems.