Multiclass Fruit Detection Using Improved YOLOv3 Algorithm

Manual interventions continue to be used in fruit-picking and billing at large-scale fruit storage facilities. Recent advances in deep in learning approaches, such as one-stage detectors like You Only Look Once (YOLO) and Single Stage Detector (SSD), as well as two-stage detectors like Faster RCNN and Mask RCNN, aim to streamline the processes involved with fruit detection and enhance efficiency. However, these frameworks continue to suffer with multi-scale objects, in terms of performance and efficiency due to large parameter sizes. These problems increase when multi-class fruits are encountered. We propose an improved version of the one-stage detector framework YOLOv3 for multi-class fruit detection. Our proposed model addresses the challenges of multi-scale object detection and detection of different fruit types in an image by incorporating CNN, bottleneck, and Spatial Pyramid Pooling Fast (SPPF) modules in the Head, Neck, and custom backbone of the YOLOv3 framework. Optimization of learnable parameters for computational efficiency is achieved by concatenating features at different feature map resolutions. The proposed model incorporates fewer layers and parameters compared to YOLOv3 and YOLOv5 models. We performed extensive testing on three datasets downloaded from Roboflow and compared them with YOLOv3 and YOLOv5 models. Our model achieved mAP50 of 0.747 on Dataset 1 comprising images of apples, bananas, and oranges whereas Dataset 2 consisting of images of apples, oranges, lemon, and Pear, achieved mAP50 of 0.981. Testing the Mineapple dataset comprising on-tree images of apples of varied sizes, achieved an accuracy of 0.643. We observe that the performance of our model beats the performance of the YOLOv3 and YOLOv5 models.

@article{Shrawne2024,
  title     = {Multiclass Fruit Detection Using Improved YOLOv3 Algorithm},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {9},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Seema C. Shrawne and Jay Sawant and Omkar Chaubal and Karan Suryawanshi and Diven Sirwani and Vijay K. Sambhe},
  doi       = {10.14569/IJACSA.2024.01509100},
  url       = {https://doi.org/10.14569/IJACSA.2024.01509100}
}