On the Joint Use of CNN and OC-PCA Classifier for Cross-Domain Potato Disease Identification

Convolutional Neural Networks (CNNs) are powerful tools for image classification when performed on the same dataset. However, when trying to use the trained CNN to classify images on unseen datasets, the performance is decreased during testing. This paper proposes an approach that integrates the One-Class Principal Component Analysis (OC-PCA) classifier with CNN-extracted features to create a flexible identification system. A comparative study between CNN and OC-PCA is conducted using various datasets. The combined system shows notable performance in handling varying unseen datasets, eliminating the need for architectural tuning and data augmentation. Extensive experiments performed across intra and inter-domains demonstrate the proposed system’s ability to achieve accuracies from 82% to 100%. The proposed approach presents an improvement of the CNN adaptability to new datasets without the need for a new architecture and data augmentation while addressing challenges linked to the diversity of plant diseases.