A picture is worth a thousand words and that’s certainly true for quality and food safety inspection. In seafood processing plants, quality control inspectors take a lot of pictures: of raw material, of semi-finished and finished products, of packaging labels, of canning seams, of sanitary conditions, of shipment loading. If there is quality or food safety noncompliance, then someone likely snapped a picture of it.

For this reason, I believe that AI-enabled smart cameras are going to revolutionize visual inspection in the seafood processing sector. It’s also why our company is investing heavily in our innovative TallyVision technology.  

Traditional visual inspection relies on human inspectors to identify defects, contamination, or irregularities in seafood products. However, these methods are often labor-intensive, inconsistent, and prone to human error and bias. Enter computer vision and smart cameras—a technological advancement that has the potential to transform seafood processing by enabling automated visual inspection, real-time monitoring, and data-driven decision-making.

In a tuna cannery, for example, instead of randomly sampling hundreds of cans per day for inspection, a smart camera can inspect each one at the dizzying speed of 4 cans per second. AI is faster, better and cheaper than human inspection. Here’s how it works:

How Computer Vision Works

Computer vision is a branch of artificial intelligence (AI) that enables machines to interpret and analyze visual information from the world, much like human vision. However, unlike humans, machines convert each pixel in an image into a number and then use neural networks and other deep learning algorithms—also known as AI “models”—to recognize patterns in those numbers.

By leveraging deep learning algorithms, smart cameras, and edge computing, seafood manufacturing plants can now automatically inspect products with speed, accuracy, and scalability unlike ever before.

Training AI models for seafood inspection involves several steps to ensure high accuracy and reliability. These include:

1. Data Collection and Annotation

A diverse dataset of seafood images is collected under different conditions, angles, and backgrounds. Each image is labeled or annotated to indicate the thing being inspected such as a defect, species, size, shape, colour, quality grade or even weight. The number of images required depends on the number of classes you want to identify—i.e. 5 species, 8 defects, 14 colour classes—and their variations.

2. Model Selection

Various AI models can be used in seafood inspection depending on the use case. The types of models include image classification, object or anomaly detection, depth or weight estimation, optical character recognition (for reading text), image segmentation and edge detection. For example, if you are inspecting a bruise on a fish fillet or dent in a can, you’ll likely use an object detection model.

3. Model Training

Using powerful cloud computing, AI models are trained with thousands of annotated images. The model learns to recognize patterns and classify seafood products based on what needs to be inspected. It can take several days to process the thousands of images and train the AI.

4. Testing and Evaluation

Typically, the AI model is only trained on 70 to 80 percent of the images collected. Once the AI is trained, you use the remaining images to test and evaluate the model for accuracy, precision and recall—which are key performance indicators for computer vision models.

Recall is the most important AI metric for food safety. Coincidentally, the name also describes what happens when defective products are pulled or “recalled” from grocery store shelves. In computer vision, however, recall reflects the idea of how well the AI model recalls or identifies all actual positive cases. Here’s an example: there are 4 cans of tuna and 3 have food safety defects. If the AI model only identifies 2, the recall is 2 of 3 defective products, or 67 percent. When it comes to food safety, you want the model to be close to 100 percent in “recalling” hazardous defects. False negatives are costly.

5. Deployment and Continuous Learning

The trained AI model is next integrated into a smart camera and deployed typically over a conveyor belt or perhaps an inspection table. A live viewer can be used to monitor inspection in real-time. If errors or biases are noticed in the inspection, more images can be collected, and the model can be retrained to remove the error or bias. AI models can continuously learn from new data, improving its accuracy over time.

6. Automation

Smart cameras can ultimately be integrated into automation equipment through a programmable logic controller or PLC. Basically, the camera sends a signal to a mechanical or pneumatic sorter to remove defective products from the production line, saving labour costs and significantly improving quality and food safety.

Use Cases in Seafood Manufacturing

Smart cameras are a remarkably scalable and adaptable technology throughout the manufacturing process, from raw material receiving to finished product inspection. Below, I’ve provided some examples of the type of visual inspection that could occur and the type of AI model to be used. This list is not exhaustive, and I challenge you to look through your own production processes for even more examples.

Raw Material: Sizing, Grading, Estimating Weight, Identifying Species

Semi-Finished Products: Defect detection, product grade, colour classification

Finished Products: Seal Defects, Seaming Defects, Mislabelling, Packaging Defects

Benefits of Using Smart Cameras in Seafood Processing

There are several benefits to AI. However, the most extraordinary is that computer vision can occasionally perform tasks that simply can’t be done by humans. For example, we’ve recently trained an AI model to estimate the percentage of flaked versus chunked tuna in a can. The AI model has an error of +/- 1.5%. Since it almost equally over- and under-estimates, the error after averaging 500 cans is only a fraction of a precent. The common advantages of using AI-enabled smart cameras include:

1. Increased Accuracy and Consistency

AI models eliminate human errors, bias and inconsistency related to physical fatigue in workers, ensuring that every batch meets the required standards.

2. Faster Processing Speeds

Smart cameras can inspect thousands of products per hour, reducing bottlenecks in the processing line. QC methodology changes from random sampling to continuous, 100-percent inspection.

3. Cost Savings and Labor Efficiency

Smart cameras reduce reliance on manual inspectors, allowing workers to focus on higher-value tasks such as process improvement. Cameras can also be connected to automated equipment for sorting.

4. Real-Time Monitoring and Decision Making

AI-powered systems provide instant feedback to plant managers, allowing for real-time adjustments and preventing costly defects from going unnoticed.

5. Undiscovered, Data-Driven Insights

Smart cameras and AI systems can collect and analyze hundreds of thousands—even millions—of datapoints each day, helping manufacturers optimize processing techniques, improve yield and reduce waste.

Conclusion

We’re only beginning to unlock the value of computer vision for visual inspection in seafood manufacturing. We’ve already seen unexpected use cases. We recently developed an OEE (Overall Equipment Effectiveness) solution using our TallyVision smart camera to count cans and track downtimes in tuna canneries. There are many OEE solutions on the market that depend on physically connecting sensors to automated equipment. Our TallyVision OEE solution, however, is the first to solely use smart cameras.

Please reach out to me at eric [at] this.fish if you’d like to partner with us to build visual inspection solutions using our TallyVision platform.