A weakly-supervised approach for flower/fruit counting in apple orchards

• A weakly-supervised regression-based deep learning algorithm. • Simplified approach for apple flowers and fruits counting without precise detection and segmentation. • Implementation of visualization techniques to investigate activated regions and features contributing to the count. • Evaluation on datasets acquired from unstructured commercial orchard environment. Flower and fruit count is a critical metric in developing crop-load management and harvesting strategies during flower/fruit development and harvest seasons. Growers currently rely on their prior experience and/or manual count in sample areas/trees to estimate the number of flowers/fruits in orchards. In this work, we propose a simplified yet robust deep learning-based weakly-supervised flower/fruit Counting Network (CountNet) and investigate its accuracy in commercial orchard images. Unlike detection-based counting methods, which require individual object detection, CountNet learns from image-level annotation with the number of objects (flowers or fruits) as input without explicitly specifying the object’s signature and location. Experiments were conducted in images acquired in an unstructured commercial orchard environment. Results showed a minimum Mean Absolute Error (MAE)/Root Mean Square Error (RMSE) of 12.0/18.4 and 2.9/4.3 for the apple flower and fruit dataset respectively. Activated region/feature visualization techniques revealed that CountNet is looking into different apple flower/fruit edges and features to make the count decisions. The results are promising in simplifying the automated methods for flower/fruit counting which can lead to reduced manual counting in the field, manual image annotation, and computational complexity and memory requirement of the object counting system.