ARTD-Net: Anchor-Free Based Recyclable Trash Detection Net Using Edgeless Module.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Science

Background:

• Household waste management faces challenges due to increasing volumes and the difficulty of recycling without separate collection.
• Manual waste separation is costly and time-consuming, necessitating automated solutions.
• Existing datasets lack sufficient diversity and resolution for complex, real-world waste scenarios.

Purpose of the Study:

• To develop and evaluate deep learning models for automatic detection and classification of recyclable household waste.
• To address the limitations of existing datasets by creating a new, comprehensive dataset.
• To improve the efficiency and accuracy of waste recycling processes through advanced computer vision techniques.

Main Methods:

• Proposed two Anchor-free-based Recyclable Trash Detection Networks (ARTD-Net1 and ARTD-Net2) utilizing edgeless modules.
• ARTD-Net1: A one-stage model with centralized and multi-scale feature extraction and an edge-weighted prediction module.
• ARTD-Net2: A multi-stage model employing a region proposal network and RoIAlign for sequential classification and regression.
• Introduced a new dataset with high-resolution images of various overlapped waste types.

Main Results:

• Both ARTD-Net1 and ARTD-Net2 demonstrated competitive performance in mean average precision and F1 score compared to existing models.
• ARTD-Net2 achieved higher accuracy, while ARTD-Net1 offered faster processing speeds.
• The new dataset significantly improved waste detection performance, especially for complex arrangements of overlapped waste.

Conclusions:

• The proposed ARTD-Net models offer effective solutions for automatic recyclable trash detection.
• The novel dataset enhances the robustness and applicability of deep learning models in real-world waste management scenarios.
• Automated systems using deep learning are crucial for efficient and accurate waste recycling.