Balancing Accuracy and Computational Efficiency: A Faster R-CNN with Foreground-Background Segmentation-Based Spatial...

Area of Science:

• Computer Vision
• Machine Learning
• Botany

Background:

• Wild plant recognition technology faces challenges in balancing model complexity, accuracy, and data processing on resource-constrained hardware.
• Non-invasive computer vision methods are crucial for avoiding damage to fragile wild plants during identification.

Purpose of the Study:

• To propose an improved lightweight Faster R-CNN architecture (ULS-FRCN) for efficient wild plant recognition.
• To address the critical issue of balancing model complexity, recognition accuracy, and data processing difficulty on resource-constrained hardware.

Main Methods:

• Developed ULS-FRCN featuring a Light Bottleneck module (depthwise separable convolution) and a Split SAM lightweight spatial attention mechanism.
• Implemented unsharp masking preprocessing to enhance model performance and reduce data processing costs.
• Validated ULS-FRCN on the PlantCLEF 2015 dataset with five wild plant species.

Main Results:

• ULS-FRCN significantly outperformed the baseline model, showing improvements of 12.77% in mAP, 0.01 in mean F1 score, and 9.07% in mean recall.
• The lightweight design and attention mechanism reduced training parameters, improved inference speed, and enhanced computational efficiency compared to the original Faster R-CNN.
• Achieved superior performance in terms of mAP, mean F1 score, and mean recall.

Conclusions:

• ULS-FRCN offers an effective solution for wild plant recognition on resource-constrained devices, such as those used in forestry.
• The proposed architecture enables efficient plant identification and management without reliance on high-performance servers.
• Demonstrated the suitability of the ULS-FRCN approach for practical deployment in ecological monitoring and conservation.