Growing cranberries requires regular inspection of the plants to ensure that there is nothing impeding their growth development. Our project aims to train a deep convolutional network to automatically detect and classify images of cranberries by growth stage, and allow for rapid, automated identification on new images. Challenges in accurately identifying growth stage include lighting conditions and detecting individual plants within the dense groups of cranberries captured in each image, in addition to correctly discriminating between stages. Since the initial project proposal, our scope has increased to include training a second model to compare to our originally planned model, to examine the relative advantages to each approach.

This review required heaps of understanding of cranberry crop physiology, image processing and computer vision techniques. The main challenge with this study is that cranberries are a perennial crop and morphologically look very small. Calculating developmental stages should consider the carryover effect from the previous year since it's a perennial crop.

Our project focused on two different Convolutional Neural Network (CNN) architectures: ResNet and YOLO. The ResNet model was included after experiencing how time-consuming manual annotation of the cranberry training images was. Our intent was to compare the accuracy of YOLO, which requires annotation of images to mark objects with a bounding box, to that of ResNet, which can be trained with whole unannotated images. These two models were chosen as they are top-of-the-line models known for speed and accuracy in object detection.