PlantSR: Super-Resolution Improves Object Detection in Plant Images

Area of Science:

• Computer Vision
• Agricultural Technology

Background:

• Deep learning models for plant image object detection are sensitive to input image quality.
• Low-resolution images hinder the performance of these computer vision models.

Purpose of the Study:

• To investigate the effectiveness of super-resolution technology in enhancing plant image object detection.
• To develop and evaluate a super-resolution model specifically for plant images.

Main Methods:

• A new dataset, PlantSR, with 1030 high-resolution plant images was created.
• A novel super-resolution model was developed and benchmarked against existing methods.
• The impact of super-resolution pre-processing on apple and soybean seed counting was assessed using YOLOv7 and P2PNet-Soy models.

Main Results:

• The developed super-resolution model outperformed state-of-the-art models on the PlantSR dataset.
• Super-resolution pre-processing significantly reduced mean absolute error in apple counting (from 13.085 to 5.71) and soybean seed counting (from 19.159 to 15.085).

Conclusions:

• Super-resolution technology offers substantial improvements for plant image object detection.
• This approach enhances the accuracy of detecting and counting specific plants, with applications in agricultural monitoring and analysis.