Efficient deep learning for mobile weed classification

Agricultural production across regions faces several challenges in cultivation, including climate variability, pests, diseases, and weeds. Deep neural network models have proven efficient in addressing some of these challenges, for example, in identifying and classifying weeds. A practical way for a farmer to monitor the health of their crop is to take a picture of a plant using their mobile device’s camera and run an application on the device that analyzes the image using a neural network. If it identifies the invasive plant, the application can determine the weed type in the image. The problem with developing this type of application is that many deep learning models require significant computational power, making them unsuitable for mobile devices. To mitigate this problem, we developed a pipeline for evaluating deep learning models suitable for weed classification. Our solution uses a set of metrics to assess both computational efficiency and model performance. Furthermore, we use our pipeline to generate an optimized model that performs inference on mobile devices with performance comparable to that of traditional deep learning models.