Ai, Industry-4-0 6 October 2022

Detection of defects in fish loins using machine vision and deep learning

detection of defects in fish
Share post

Machine vision technology assisted by deep learning is an important ally for fish processing and distribution factories that makes it possible to inspect 100% of the production to ensure high standards of quality and food safety of the product that finally reaches the consumer’s table.

The new Visum DeepSight Loins™ system from IRIS Technology is a machine vision system designed for the detection of physical surface defects in fresh and frozen fish loins that makes it possible to automate the inspection of loins, quantify, classify and reject non-conformities to ensure superior quality of the final product.

Machine Vision and Deep Learning

While traditional computer vision systems learn to classify and recognize features from a set of historical images in order to correctly predict and classify new ones, deep learning neural networks are able to learn features from pixels (individual and group) and have an input layer (the raw image), a series of intermediate layers that are interconnected to simulate how a biological brain works, and an output layer that provides classification/prediction. Deep learning neural networks are especially good at learning complex features and segmenting an image at different levels of abstraction (edges, different colors, shapes, objects), including noise and probabilistic information.

Traditional machine vision that does not use this approach typically processes images but does not learn from the data, such as thermal imaging cameras, motion detection sensors, light intensity sensors, among others.

Detection of defects in fresh and frozen fish loins

The DeepSight Loins™ system is capable of detecting numerous defects in fish loins such as bruises, blood stains, gapping (i.e. openings or tears in the musculature), skin remnants, superficial bones or other superficial foreign bodies that may reach the processing line. It also has built-in color measurement functionality under international CIELAB or L*a*b* standards, which is important as a quality parameter both on the surface and in relation to the freshness of the fish.

DeepSight Loins™ has a high IP protection for easy cleaning of the line and has a built-in anti-reflective and anti-humidity system that allows it to operate normally on both fresh and frozen fish loins.

Usability, Operation and Communication

The Visum DeepSight Loins™ system incorporates 2 user levels: “Administrator” for modifying settings, working mode, adjusting rejection sensitivity or taking references and “Operator” for automatic operation mode of the device.

The system is complemented by a trap door rejection that allows the ejection of non-conforming units for reprocessing or control by the operators.

The information and results of the analysis, such as the quantification of defects and rejects by class, lot information and the quantity of products inspected, can be viewed on the built-in computer module, on a computer connected to the network or on the plant’s own information management system. In addition, automatically generated reports can be exported in different formats.

The sensitivity adjustment functionality is an essential tool for calibrating the level of rejection of the device in the event of certain defects and thus regulating the system’s operating performance without causing any inconvenience to the line’s production capacity.

For more information about the device and inquiries write to

By IRIS Technology Solutions
Ai, Ispe, Webinar 19 November 2020

IRIS presents the Predictive System build for mAbxience at the “ISPE 4.0 AI applications” webinar

Share post

On november 25, the workgroup of ‘ISPE 4.0: AI applications to pharmaceutical processes’ organized a webinar about “Digitization and advanced analytics in pharmaceutical”, where Dr. David F. Nettleton, Senior Data Mining Analyst at IRIS, presented, together with Francisco Javier Rodríguez, Engineering Director at Genhelix (mAbxience), how IRIS developed a successful predictive system to anticipate possible events or anomalies in their plant.

Predictive system

In their presentation, both companies, mAbxience and IRIS, explained both, how they started a collaboration a few months ago with the ambitious purpose of creating a Predictive System that could help to anticipate events or anomalies in one of the  mAbxience plants, and how they successfully achieved their objective.

For this Predictive System, mAbxience collected over 115 million values from the critical process variables of the mAbxience plant, and 10 thousand messages from their alarms system, that were filtered with Big Data pre-processing techniques by IRIS, to clean, consolidate, cluster and normalize this huge amount of data for their future analysis.

Predictive system

Example of clustered values in the Database (showing normal cases in blue and anomalies in red)

By using HPC (High Performance Computing), a  Distributed Processing Architecture and Artificial Intelligence techniques like Machine Learning, IRIS build the predictive models       that were able to identify the key trends in the data that could lead to an undesired events or anomaly.

Predictive system

Simplified example of the likelihood of having an event in three different sensors.

This successful case between mAbxience and IRIS is an excellent example of the work done in the ISPE Spain workgroup, that was created with the aim of meeting people from the pharmaceutical and chemical sectors whose work is focused on Industry 4.0 and the implementation of PAT (Process Analytical Technologies) in the laboratory.

By Lorena Vázquez