A comparable study: Intrinsic difficulties of practical plant diagnosis from wide-angle images

Practical automated plant disease detection and diagnosis for wide-angle images (i.e. in-field images contain multiple leaves from fixed-position camera) is a very important application for large-scale farms management, ensuring the global food security. However, developing the automated disease diagnosis systems is often difficult because labeling a reliable disease wide-angle dataset from actual field is very laborious. In addition, the potential similarities between the training data and test data leads to a serious model overfitting problem. In this paper, we investigate changes in performance when applying disease diagnosis systems to different scenarios of wide-angle cucumber test data captured in real farms and propose a preferable diagnostic strategy. We show that the leading object recognition techniques such as SSD and Faster R-CNN achieve excellent end-to-end disease diagnostic performance on only the test dataset which is collected from the same population as the training dataset (81.5% - 84.1% F1-score for diagnosed disease cases), but it seriously deteriorates on the completely different test dataset (4.4 - 6.2% of F1-score). In contrast, the two-stage systems which have independent leaf detection and leaf diagnosis model attain a promising disease diagnostic performance with more than 6 times higher than the end-to-end systems (33.4 - 38.9% of F1-score) on the unseen target dataset. We also confirmed its efficiency from visual assessment, concluding that the two-stage models are suitable and a reasonable choice for the practical application.