Solution: Computer Vision-based Contamination Detection

Benefits

Increased Efficiency 

An AI-based timber detection system can identify and isolate timber contamination in real time with high accuracy. This helps in maintaining a smooth and uninterrupted operation, reducing the time spent on manual inspection or dealing with equipment blockages caused by large timber pieces.

Improved Product Quality

The system can drastically reduce the amount of timber contamination that reaches the alumina refining stage. This minimises the risk of process contamination and improves the overall quality of the end product, bauxite ore and alumina, by ensuring that it is free from organic timber matter.

Enhanced Safety  

By automatically identifying and removing timber pieces, the system reduces the risk of equipment damage and potential safety hazards associated with blockages in crushers and chutes and with the manual retrieval of timber. It also minimises human exposure to the operational area, providing a safer working environment.

Technology

Leveraging YOLO's real-time object detection capabilities, it's possible to develop a system that can rapidly and accurately identify instances of timber in the extracted ore. With its robust balance of speed and accuracy, YOLO can operate on conveyor belts or similar systems to identify and localise timber contamination in real-time, allowing for prompt removal and reducing the likelihood of process downtime and downstream contamination. Its versatility and adaptability mean our team could train the network to recognise different types of wood and sizes of debris, making it a valuable tool for enhancing the efficiency and safety of mining operations.

The YOLO framework, noted for its effective balance of speed and accuracy, has seen multiple iterations, each building upon the previous ones to address limitations and enhance performance. This paper explores the key innovations, differences, and improvements across each version, from the original YOLOv1 to the latest YOLOv8. These improvements span various aspects including network design, loss function modifications, anchor box adaptations, and input resolution scaling. The paper also discusses the trade-offs between speed and accuracy that have marked the framework's development, underlining the importance of contextual and application-specific requirements when choosing the most suitable model.

DETRv2 benefits from being an end-to-end transformer-based model. It eliminates the need for anchor boxes and non-maximum suppression (NMS). This simplifies the pipeline and reduces error sources. DETRv2 leverages transformers to encode global relationships between objects in the image, which helps with more accurate detection of objects in complex scenes. Transformers allow the model to focus on both local and global interactions simultaneously. DETRv2 with deformable attention mechanisms can handle small objects and occlusions better, as it dynamically focuses on the most relevant parts of the feature maps, leading to better handling of challenging objects or cluttered environments. It generally handles larger images and resolutions better due to the attention mechanism’s ability to scale to larger contexts. DETRv2's transformer architecture is inherently more flexible and extendable for other tasks like segmentation and panoptic tasks. DETR models are well-suited to multi-task setups without significant architectural changes.