FruitNet: Lightweight CNN for High-Throughput Image-Based Fruit Yield Estimation

Estimation offrait yield is crucial for agricultural practices to be optimized and secure food supply. The current methods of estimating yield are labour intensive and inaccurate that resulted in the development of more advanced technological solutions. Innovations in this work include the most lightweight Convolutional Neural Network (CNN) named FruitNet proposed for achieving high throughput and image based estimation offrait yield. Based on state of art deep learning techniques, FruitNet predicts fruit yield from the image offrait bearing plants with accurancy and efficiency. We propose a model that is computationally efficient, and feasible to deploy in unconstrained, constrained, and resource constrained environments, namely, in small farms, remote agricultural areas and many more other such environments. A streamlined architectural design of FruitNet with minimal computational load but maintaining high prediction accuracy is devised. Therefore, in order to ensure that the model is robust to different scenarios the model is trained on a robust dataset involving fruit of different variety, growth stage and under different environmental conditions. Empirical evaluations confirm that FruitNet matches the accuracy of more complex models at the cost of much less inference time and resource consumption. By making the model lightweight, fast and easy to deploy on edge devices, the model can serve as aid for farmers in real time to estimate yield, and make decisions. Moreover, FruitNet's high throughput with images enables it to be useful for high throughput agricultural operations.