Research on Apple Recognition Algorithm in Complex Orchard Environment Based on Deep Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Traditional fruit recognition algorithms struggle with accuracy, real-time processing, and robustness in complex orchard environments.
• Low accuracy and poor performance hinder automated harvesting and yield estimation.

Purpose of the Study:

• To develop an improved deep learning-based fruit recognition algorithm for complex orchard environments.
• To enhance recognition accuracy, real-time detection speed, and robustness compared to existing methods.

Main Methods:

• Integrated residual module with Cross Stage Partial Network (CSP Net) to optimize performance and reduce computational load.
• Incorporated Spatial Pyramid Pooling (SPP) module into YOLOv5 to improve detection of small fruit targets.
• Replaced Non-Maximum Suppression (NMS) with Soft NMS to better handle overlapping fruits.
• Constructed a joint loss function using focal and CIoU loss for improved accuracy.

Main Results:

• Achieved a Mean Average Precision (MAP) of 96.3%, an improvement of 3.8% over the original model.
• Reached an F1 score of 91.8%, a 3.8% increase compared to the baseline model.
• Attained an average detection speed of 27.8 frames/s on GPU, 5.6 frames/s faster than the original model.

Conclusions:

• The improved deep learning algorithm demonstrates excellent detection accuracy, robustness, and real-time performance in complex orchard settings.
• This method provides significant improvements over existing techniques like Faster RCNN and RetinaNet.
• Offers valuable insights for accurate fruit recognition challenges in agriculture.