The project began with the analysis of a vast image database, a process that took nearly 6 months to appropriately classify. The acquired training material served as the foundation for developing multiple classification models using Machine Learning, particularly Deep Learning techniques.

Two Overlapping Neural Networks

We simultaneously employed two neural networks because deep learning has its limitations. FIRECRUX models for AI was trained using two datasets: one comprising images of pristine, uncontaminated products, and the other containing images of products mixed with contaminants. This dual dataset strategy allows us to address three key scenarios:

1. Verification of a clean production line – in this case, the classification model trained on clean product images exhibits the highest score.
2. Detection of contamination within the scanned area – the scoring is reversed to identify the presence of contaminants.
3. Ensuring manufacturer safety in unusual situations involving impurities outside the tested spectrum or intentional sabotage – in such instances, the models produce similar scores, signaling a suspicious situation. For safety reasons, these situations are flagged and rejected for investment by our staff.

Hardware Requirements

System that has been developed is based on the FIRECRUX engine and can be operated on GPU, CPU, or VPU processors. In the current version at Milarex, it runs on a GPU processor using the nVidia Jetson TX2 Developer Kit (945-82771-0005-000). The CPU version is embedded on Intel i7 and i9 processors, while the VPU version utilizes Intel Movidius.

Furthermore, the software has been seamlessly integrated with X-ray lamps, including the SAXG monobloc x-ray generators from Spellman and a line scan detector from Detection Technology.

FP Reduction and Effectiveness Testing

The previous solution utilized by Milarex, developed by a Canadian company, led to significant downtime due to the occurrence of false positives (FP), averaging nearly 2 hours per day. During the initial phase of our solution’s development, around 2017, we placed considerable emphasis on addressing this FP issue. Through rigorous testing, both on quality control testers and in real-world contamination scenarios (such as detecting broken knives from portioning machines, identifiable medical plasters, or glass fragments), we successfully reduced the FP rate from approximately 20% to less than 1%.

Subsequently, we resolved this problem and integrated the option to mark FP (False Positives), NP (Non-Presence), and TP (True Positives) within the X-ray scanner management system.

We rely on machine-generated data, specifically the number of scans, detections, and saved confirmation photos, to monitor the system’s performance.

Real-time Contamination Detection

The completed system also enables on-the-fly scanning. Our software has been designed to accommodate two types of machines: cyclic and continuous scanning.

The cyclic scanning method proved sufficient for machines handling 9/12/16 pieces of salmon, as the production line’s manual worker efficiency was the bottleneck in these scenarios.

In 2021, a continuous scanning solution was developed. Instead of having 9/12/16 trays go into the bin upon contamination detection in the cyclic machine, the fish are now seamlessly redirected to the real-time scanning machine without the need for a ‘stop-start’ action.

Production Deployment and Licensing

The solution for a first machine was implemented in 2019 and an additional 13 machines were gradually added between 2019 and 2022. Subsequently, development efforts continued until November 2022 and a local data cloud was established.

The final version of the system achieved two significant milestones:

• A 61% increase in contamination detection efficiency,
• A 66% increase in production line speed.

In Nov 2022 licensing and maintenance agreements were signed and currently Greenlogic is solely dedicated to optimizing classification models to maintain a high level of contamination detection accuracy.