Pear and apple recognition using deep learning and mobile

Abstract

Apple and pear are among the most widely grown and economically important fruit species worldwide and in Latvia. In turn, scab diseases caused by ascomycetous fungi Venturiainaequalis and V. pyrina, are economically the most important diseases worldwide. Durable plant resistance has been widely regarded as the preferred disease limitation method due to environmental and food safety concerns. Whereas in cases where the use of pesticides cannot be avoided, their applications should be more precise, more targeted and reduced substantially. One way how to realize it is the smart and precision horticulture that can greatly increase the effectiveness of pesticides and use them more selectively. The smart and precision horticulture relies heavily on new technologies and digitalization, including sensing technologies, software applications, communication systems, telematics and positioning technologies, hardware and software systems, data analytics solutions, as well as knowledge linking biological information to data technologies. The aim of our project - development and implementation of mobile application with deep learning system for early identification and evaluation of apple and pear scab. The specific of project - the image processing must be completed by a mobile device without image upload into GPU cluster. This research presents the comparison of deep learning architectures developed for mobile devices (MobileNet and MobileNetV2). The classification precision and speed of neural networks are compared using open dataset”Fruits-360”. The results are applicable to develop transfer learning and domain adaptation solutions. Meanwhile, decomposition into many simple subtasks can reduce required device resources to complete complex analysis using mobiles, as well as to create trustworthy AI model. The model of MobileNetV2 showed the best results: total accuracy 0.998, Cohen's Kappa 0.991 and latency 212ms/step.