Disease classification from potato leaf images (MS) : an empirical study on images captured in unconstrained environments

Abstract

The aim of this work is to classify potato plant diseases from images of potato leaves, captured in the field. Such images are characterised by a rich background consisting of multiple leaves, weeds, ground patches etc. In addition, there are possibilities of a wide range of illumination variation and defocus. All these make discriminative feature learning an exceedingly difficult task. The diseases being considered are early and late blight which predominantly affect the leaves and present as discolouration of varying sizes. This work focuses on classifying a given leaf image as healthy or affected by early or late blight. An initial classi cation study reveals many anomalies, the most signifi cant one being that classi cation accuracy does not improve despite using the segmented fore- ground leaf. In order to explain this, a detailed analysis of learnt weights and the features is carried out. The latter was done using visualisation methods to identify the regions in an input image which contributed towards the features. The analysis shows the impact of dataset bias, and domain adaptation and generalisation seems to be possible directions for building an effective classi er. However, the analysis also reveals a more severe issue that most architectures do not learn features from relevant regions (disease spots and healthy leaf) in the image due to the fact that the percentage of pixels representing disease spots are too less. Also, the appearance of these regions vary a lot with illumination and defocus making them look like healthy regions. This finding questions most of the reported results of similar works. In addition, it also points to the fact that it is essential to have either an attention based solution or a solution which is region based. A region based solution is explored using faster region based convolutional neural network (faster R-CNN) as region of interest (ROI) detector followed by a classi er. A ground-truth overlap index of 91.85% and an overlap recall of 83.06% show that the relevant regions are extracted. This in turn improves the classi cation accuracy remarkably. In order to test the generalisability and hence the reliability of a model, the notion of cross-testing is introduced wherein a classi er trained with one dataset is tested using a different dataset, the difference being in the nature of backgrounds. Unlike the other models, the ROI-based classi er shows exceptionally good classi -cation accuracy during cross-testing. While the conventional CNN-based classi ers gave an average cross-testing accuracy around 50%, the faster R-CNN-based classi er exhibited a cross-testing accuracy of 84.76% on lab-prepared dataset. If not for the low resolution images of the lab-prepared dataset, this accuracy could have been higher. This classi er achieved 96.95% testing accuracy for the in- field dataset, on which the model was trained. Following the observations and analysis of all the results, we come up with a set of clear directions to create an image dataset, which can lead to a reliable leaf image based plant disease classi er.