The visual and tactile examination of plant leaves is a standard method for identifying disease in crops and horticultural products. However, such an approach can be highly subjective and is dependent on the skills of the examiners. Writing in the International Journal of Computational Vision and Robotics, a team from Egypt describes a new approach to plant leaf disease detection using deep learning on a mobile device. The team’s tests against a standard database of diseased leaf images showed their system to be capable of up to 98 percent diagnostic accuracy. The process is rapid and showcases the sophisticated computational power available in modern mobile phones for this kind of intensive task.
Shaheera A. Rashwan and Marwa K. Elteir of the Informatics Research Institute at the City of Scientific Research and Technological Applications in Alexandria, suggest that for busy farmers in remote regions with no immediate access to plant disease experts, a mobile application that can help them spot disease and so treat the crops in a timely manner could be vital to their ongoing agricultural viability.
The team’s approach exploits the recent evolution of computational systems and especially graphical processing units (GPUs) that allow machine learning operations to be carried out efficiently in ways that previous generations of devices simply could not match for speed. Such operations facilitate the running of tools such as convolutional neural networks, which mimic certain characteristics of brain function, and allow image recognition and related tasks to be carried out quickly. The team thus embedded image recognition of the characteristics of disease in leaves for the present research.
Despite the great speed and accuracy of disease diagnostics that the team has shown, there is still room for improvement. They highlight an issue with shadows on images and confusing backgrounds when a user takes a photo of a suspect leaf. They hope to be able to develop a pre-processing step that will reduce any problems and the inaccuracies that might arise if the acquired leaf image is not as perfect as it might be for image recognition. Fundamentally, automated light level adjustment in the image would preclude issues arising because of shadows, while a step that isolates the leaf from its background in the image and effectively removes said background would ease the whole process still further and hopefully nudge the accuracy upwards.
Rashwan, S.A. and Elteir, M.K. (2022) ‘Plant leaf disease detection using deep learning on mobile devices’, Int. J. Computational Vision and Robotics, Vol. 12, No. 2, pp.156–176.