Deployment-Optimized Deep Learning Model for Coconut Shell Quality Identification in Charcoal Production

Coconut shells are a vital raw material for charcoal briquette production, and ensuring their quality is crucial for producing high-quality charcoal products. This study develops an optimized deep learning model for detecting, classifying, and evaluating coconut shell quality using Mask R-CNN, with a specific focus on bridging the gap between theoretical model development and practical industrial implementation. Unlike previous studies that only evaluate model performance in Integrated Development Environments (IDEs), this research conducts comprehensive evaluation across both IDE and industrial web platforms. The Mask R-CNN model is modified by replacing the default ResNet-101 backbone with optimized variants including ResNet50-FPN and MobileNet-FPN to address performance degradation during deployment. A dataset comprising 1,611 coconut shell images with 8,922 annotated objects across four quality classes was created for training. Experimental results demonstrate that MobileNet-FPN achieves optimal balance between accuracy and computational efficiency, with 96% mAP@0.50 on training and 84% mAP@0.50 on validation. Class-wise analysis reveals "Wet and Fibrous" as the most challenging class (74.8% AP@0.50) due to feature overlap with parent classes and class imbalance effect, while "Clean and Dry" achieves highest performance (91.2% AP@0.50) due to distinctive visual characteristics. Statistical analysis confirms significant performance differences between architectures (p < 0.05). Deployment evaluation reveals MobileNet-FPN consistently achieves average inference time of 4.79 seconds on web platform with lowest variance (SD = ±2.41s), suitable for industrial quality control applications. The developed Flask-based web application enables charcoal briquette producers to evaluate raw material quality efficiently, demonstrating the importance of deployment-aware optimization for practical industrial application.