Improving Vision-Based Detection of Fruits in a Camouflaged Environment with Deep Neural Networks

Abstract

In agriculture, camouflage, i.e. fruits are occluded due to high color similarity to its environment, is one of the problems of fruit detection and counting for yield estimation. In this paper, we developed a system that automatically detects and counts fruits from images in an unstructured environment with heavy occlusion and high color similarity. We built two datasets by capturing the natural settings of the fruit and fine-tuned various deep convolutional neural networks with region proposal network for a very challenging task of detecting heavily occluded fruits with high color similarity against the background. The evaluation results show an average precision of 0.922 for chili pepper dataset and 0.966 for bell pepper dataset. The result of this network can be integrated into an application for efficient management of agricultural farms, thereby contributing to precision agriculture.