A deep-level region-based visual representation architecture for detecting strawberry flowers in an outdoor field

Abstract

An accurate and robust strawberry flower representation and detection scheme is a key step to enable the reliable forecasting of fruit yield for use in precision agricultural applications. A state-of-the-art deep-level object detection framework which processes images through several layers using a region-based convolutional neural network (R-CNN) was developed to visually represent the instances of strawberry flowers in outdoor fields and improve the detection accuracy. A modified version of the visual geometry group 19 (VGG19) architecture, which had 47 layers, was used to represent the multiple scales of strawberry flower image features. The networks were trained entirely on 400 strawberry flower images and tested on another 100 images. Different region-based object detection methods, including the R-CNN, Fast R-CNN and Faster R-CNN, were used to represent the strawberry flower instances. The Faster R-CNN model achieved a better performance than the R-CNN and Fast R-CNN in detecting the instances and had a lower execution time. The detection accuracy of the Faster R-CNN model was 86.1%, which was higher than those of the R-CNN and Fast R-CNN models (63.4% and 76.7%, respectively). The experimental results showed the effectiveness of the deep-level Faster R-CNN framework for representing the strawberry flower instances under various camera view-points, different distances to flowers, overlaps, complex background illumination, blur, etc. The system developed for automatic and accurate strawberry flower detection provides an important and significant solution that enables subsequent applications to estimate the strawberry yield in outdoor fields.