Comparative analysis of deep learning frameworks for automated fruit detection in smart agriculture

Recently, deep learning-based object detection approaches have come into prominence to automatically determine the fruits on trees for smart agriculture. However, the suitability of various frameworks to different fruit types has not been adequately analyzed in such studies. Taking this into consideration, in this thesis, the performance of different deep learning frameworks for object detection in detecting different fruit types has been comprehensively analyzed. For this purpose, RetinaNet, YOLOv3, and YOLOv4, which are among the highest performing deep learning frameworks for object detection according to the current literature, and 5 different datasets for mango, almond, grape, and apple images were used. After a through experimental analysis, it has been observed that RetinaNet is the best framework in terms of detection accuracy and training and detection time. Also, it has been found that the factors such as atmospheric conditions, properties of the object, and time of day while constituting the image datasets have great impact on the detection performance. Finally, it has been noted that documenting the conditions of the experiments facilitates the reproducibility of the experiments and comparison between different frameworks and datasets.