Comparative Analysis of Banana Detection Models

Abstract

This study aims to compare and evaluate the performance of banana detection models utilizing deep learning techniques and the Darknet algorithm. The objective is to identify the most effective approach for accurately detecting bananas in various real- world scenarios. The analysis involves training and testing multiple models using different datasets and evaluating their performance based on precision, recall, and overall accuracy. The results provide valuable insights into the strengths and weaknesses of each approach, enabling researchers and practitioners to make informed decisions when implementing banana detection systems. To detect banana objects, several convolutional neural network (CNN) models were used, including MobileNetV2, YOLOv3-Nano, YOLO Fastest 1.1, YOLOv3-tiny-PRN, YOLOv4-tiny, YOLOv7, and DenseNet121-YOLOv3. The training process utilizes the Darknet algorithm to facilitate the identification of banana types/classes captured by a camera, resulting in an MP4 film file. In this research, various experiments were carried out using different CNN models. However, these six models achieve optimal accuracy above 80%. Among them, the YOLOv7 model shows the highest average accuracy (MAP) at 100%, followed by the small model YOLOv4 at 92%. Meanwhile, for performance measurements, the accuracy of the YOLOv4-tiny model was 87%, followed by the YOLOv7 model at 84%. In the banana fruit experiment, several models showed very good performance, such as recognition of the Ambon, Kepok, and Emas banana classes up to 100% using the YOLOv7 and YOLOv4-tiny models. The YOLOv7 model itself can recognize other banana classes up to 100% in the Barangan, Rjbulu, Uli, and Tanduk classes. Furthermore, theYOLOv4-tiny model can recognize other banana classes, up to 90% of the Barangan, Rjbulu, Rjsereh, and Uli banana types. Thus, this experiment provides very good average accuracy results on 2 CNN models, namely YOLOv7 and YOLOv4-tiny. Future research will involve grouping pictures of bananas, which produces different image shapes, so it requires a different way to recognize them. It is hoped that this research can become a basis for further research in this field.