An Ensemble Learning Approach towards Waste Segmentation in Cluttered Environment

1. Problem Statement

The paper addresses the critical challenge of automated waste segregation in Material Recovery Facilities (MRFs). While recycling is essential for environmental sustainability, current MRFs rely heavily on manual labor, which is inefficient and prone to error.

• Core Challenge: Waste in real-world environments (conveyor belts) is highly cluttered, occluded, and deformed. Objects often overlap, lack distinct patterns, and vary significantly in size.
• Limitation of Existing Methods: Traditional object detection (bounding boxes) is insufficient for robotic grasping, which requires precise pixel-level localization. Existing segmentation models often struggle with the complexity of real-world waste streams, failing to capture fine boundaries or handle scale variations effectively in cluttered scenes.
• Goal: To develop a deep learning segmentation model capable of generating accurate predictive masks for waste items in cluttered environments to facilitate robotic sorting with minimal human intervention.

2. Methodology

The authors propose an Ensemble Learning (EL) approach that combines the strengths of two high-performing segmentation architectures.

A. Dataset

• Source: The ZeroWaste-f dataset, a publicly available, industrial-grade dataset specifically designed for waste detection and segmentation.
• Characteristics: Images are highly cluttered, containing deformed objects and overlapping items, closely mimicking real-life MRF scenarios.
• Preprocessing:
  • Resizing: Images resized from Full HD to 320 x 480 to reduce computational load.
  • Augmentation: Applied to the training set (flipping, scaling, rotating, noise addition, perspective transforms, color adjustments) to diversify data.
  • Normalization: Scaled based on ImageNet statistics (mean and standard deviation per channel).
  • Transfer Learning: Models utilize pre-trained weights from ImageNet to accelerate learning.

B. Model Architecture & Selection

The study followed a two-phase approach:

1. Baseline Evaluation: Seven state-of-the-art segmentation models were evaluated on ZeroWaste-f:
   • U-Net, U-Net++, MANet, LinkNet, FPN, PSPNet, and PAN.
   • Selection Criteria: Models were chosen based on Intersection over Union (IoU) and Dice Loss.
   • Findings: U-Net achieved the highest IoU (0.8065), excelling in boundary detection. FPN (Feature Pyramid Network) achieved the lowest Dice Loss (0.1183), excelling in handling scale variations and context.
2. Ensemble Construction (The Proposed Model):
   • Strategy: A weighted average (stacking and averaging) of the segmentation masks generated by U-Net and FPN.
   • Backbone: Both models utilize EfficientNet encoders (variants B0 through B4) instead of standard ResNet/ResNeXt backbones to reduce computational costs while maintaining feature extraction quality.
   • Variants: Five ensemble models were created (EL-0 to EL-4) corresponding to EfficientNet B0 through B4.
   • Combination Logic: The masks from U-Net (fine details) and FPN (multi-scale context) are averaged to produce a final, more robust prediction mask.

C. Training Configuration

• Framework: PyTorch (using segmentation_models_pytorch and timm).
• Hardware: NVIDIA Tesla T4 GPU (Google Colab).
• Loss Function: Dice Loss (measures overlap between predicted and target masks).
• Optimizer: Adam (Learning rate: 0.0001).
• Metrics: IoU, Dice Loss, and F1 Score.

3. Key Contributions

1. Comprehensive Benchmarking: Evaluated seven SOTA segmentation models on the challenging ZeroWaste-f dataset, identifying U-Net and FPN as the top performers for this specific domain.
2. Novel Ensemble Architecture (EL-4): Proposed a hybrid model combining U-Net and FPN via a weighted averaging mechanism. This leverages U-Net's boundary precision and FPN's scale invariance.
3. Efficiency Optimization: Integrated EfficientNet backbones (B0–B4) into the ensemble framework to create a family of models (EL-0 to EL-4) that balance computational cost with high accuracy.
4. Real-World Applicability: Validated the approach on a dataset that truly represents the "cluttered" nature of industrial waste, moving beyond idealized lab conditions.

4. Results

The ensemble approach demonstrated significant improvements over individual models:

• Performance Metrics (Test Set):
  • Baseline U-Net: IoU = 0.8065, Dice Loss = 0.2084.
  • Baseline FPN: IoU = 0.7953, Dice Loss = 0.1183.
  • Best Ensemble (EL-4):
    • IoU: 0.8306 (an improvement of ~2.4% over U-Net).
    • Dice Loss: 0.09019 (a reduction of ~23.7% compared to FPN).
• Comparison of Variants:
  • EL-4 (using EfficientNet-B4) consistently outperformed other variants (EL-0 to EL-3) across training, validation, and test sets, indicating superior generalization.
  • Higher EfficientNet variants generally yielded lower Dice Loss and higher IoU, though at a higher computational cost.
• Visual Quality: The generated masks showed high fidelity to ground truth, successfully segmenting overlapping and deformed objects where individual models might have failed.

5. Significance and Impact

• Industrial Efficiency: The proposed EL-4 model offers a viable solution for automating waste sorting in MRFs. By providing precise pixel-level masks, it enables robotic arms to grasp and sort waste more accurately than bounding-box-based systems.
• Resource Recovery: Improved segmentation accuracy leads to better raw material recovery for recycling, reducing contamination and increasing the throughput of recycling facilities.
• Scalability: The use of EfficientNet backbones ensures that high-accuracy models can be deployed on hardware with moderate computational resources, making the solution practical for real-time industrial applications.
• Future Work: The study establishes a strong foundation for reducing human intervention in waste management, contributing to the broader goals of the "Zero Waste" initiative.