An occluded cherry tomato recognition model based on improved YOLOv7

Area of Science:

• Agricultural Robotics
• Computer Vision
• Deep Learning

Background:

• Occlusion is a major challenge for automated cherry tomato harvesting robots.
• Accurate recognition of occluded tomatoes is crucial for efficient robotic picking.

Purpose of the Study:

• To develop an efficient and accurate deep convolutional neural network model for recognizing occluded cherry tomatoes.
• To improve the performance of cherry tomato picking robots in natural environments.

Main Methods:

• Constructed a cherry tomato dataset (TOSL) with varying occlusion levels.
• Modified the YOLOv7 architecture, incorporating depth-separable convolutions and coordinate attention (CA).
• Replaced SPPCSPC with a depth-separable convolutional SPPF module to preserve small target information.

Main Results:

• The proposed DSP-YOLOv7-CA model achieved an average detection accuracy (mAP) of 98.86%.
• Model parameters were reduced from 37.62MB to 33.71MB.
• Demonstrated superior performance in detecting cherry tomatoes with less than 95% occlusion.

Conclusions:

• DSP-YOLOv7-CA effectively recognizes occluded cherry tomatoes in natural settings.
• The model offers a viable solution for enhancing the precision of cherry tomato picking robots.
• Further improvements may be needed for extremely high occlusion levels (>95%).