Weed detection in blackgram fields under occlusion and class imbalance using SA-ANMS Faster R-CNN

Area of Science:

• Computer Vision
• Machine Learning
• Precision Agriculture

Background:

• Sustainable agriculture faces challenges in detecting small, overlapping weeds amidst dense crop canopies, leading to scale disparities and class imbalances.
• Existing methods struggle with extreme occlusion and scale variations common in real-world field conditions, impacting weed detection accuracy.
• A dataset from blackgram fields in India highlighted significant scale diversity, with most objects occupying less than 0.8% of the image space.

Purpose of the Study:

• To develop an enhanced Faster R-CNN framework for accurate and efficient weed detection in challenging agricultural environments.
• To systematically investigate the impact of integrating various enhancement modules (Spatial Attention, Multi-Scale Fusion, Context-Aware RoI, Shape-Aware Prediction, Adaptive Non-Maximum Suppression) on detection performance.
• To identify optimal module combinations that improve weed detection accuracy and efficiency, rather than relying solely on architectural complexity.

Main Methods:

• A Faster R-CNN framework was augmented with five modules: Spatial Attention (SA), Multi-Scale Fusion (MSF), Context-Aware RoI, Shape-Aware Prediction, and Adaptive Non-Maximum Suppression (ANMS).
• An extensive ablation study was conducted, evaluating 32 different model configurations to assess the performance impact of module combinations.
• Performance metrics included F1-score, precision, recall, mean Intersection over Union (mIoU), and inference speed (FPS).

Main Results:

• The combination of Spatial Attention (SA) and Adaptive Non-Maximum Suppression (ANMS) modules achieved the highest F1-score (0.9547), precision (0.9439), recall (0.9658), and mIoU (0.9445) with an inference speed of 4.42 FPS.
• This two-module combination outperformed the full five-module version, demonstrating that increased architectural complexity does not guarantee better performance.
• Only 12.5% of the tested configurations surpassed the baseline, indicating that excessive module stacking can degrade performance; a positive synergy was observed between SA and ANMS.

Conclusions:

• Selective combination of enhancement modules is more effective than random stacking for designing efficient weed detection systems in precision agriculture.
• The optimal combination of SA + ANMS highlights the benefit of efficiency-based design over complexity-based expansion for handling occlusion and inter-class inconsistencies.
• This research provides crucial insights for developing robust and efficient detection systems tailored for precision agriculture applications.