dc.description.abstract | Problem: Recent years have seen an accelerated growth in use of small devices and intelligent applications in agriculture as well as other sectors. Due to this increased use of intelligence in small devices, there has been an increased demand for machine learning models that can work within resource constrained environments. Driven by the success achieved by deep learning models across different sectors, there has been a natural trend of pushing deep learning models in the direction of mobile application and use in small devices such as micro-controllers. However, this has not been an easy task. This is attributed to the difference that exist between the constrained computational environment of these small devices and the intense resource requirement of deep learning networks. As a result, there has been a need to optimize the existing deep learning models for use in small devices without significant loss of accuracy.
Objective: The objective of this study was to develop an optimized deep learning model for the prediction of tomato leaf diseases that can be used in low-end small devices.
Methodology: CRISP-DM methodology was used in building the initial TensorFlow model. A ResNet – 50 model was adopted and through use of Transfer Learning custom layers were built for the research. TensorFlow-Lite framework was used for the conversion and optimization processes of the model. Float16 Quantization was adopted for the optimization of the TensorFlow Model. Performance metrics including accuracy, precision and recall were used for model evaluation.
Results: The initial TensorFlow model achieved an accuracy of 95.93% while the converted and optimized model achieved an accuracy of 93.75%. After optimization, the initial model was reduced from 245 MB to 61 MB representing a 75% reduction in memory size. The accuracy loss was minimal at only 2.18% and still within incredible accuracy range.
Conclusion: The study concluded that quantization should be adopted as a standard practice for deep learning models in Agriculture to enhance easy deployment and accelerated adoption and use of deep learning models among small scale tomato farmers. The study recommended full deployment of the model as a mobile application and also the exploration and comparison of different quantization approaches which were not covered by the study. | en_US |