Deep Learning-Based Non-Destructive Seafood Freshness Assessment: A Systematic Review of Sensing Technologies, Validation, and Deployment Readiness

The journey of a fish from the ocean to a plate is a race against time. Once caught, seafood begins to change chemically and biologically, a process driven by bacteria, the breakdown of proteins, and the oxidation of fats. These changes alter the smell, texture, and appearance of the food, eventually leading to spoilage. For decades, determining whether a fish is still safe to eat has relied on methods that are either slow, require a laboratory, or destroy the product itself. A scientist might need to cut open a fish, grind it up, and run chemical tests to measure the buildup of specific compounds like volatile nitrogen or the count of living bacteria. While accurate, this approach means the very fish being tested can no longer be sold. The food industry and consumers alike have long sought a way to peek inside the quality of seafood without breaking the seal or ruining the catch.

A new systematic review published in October 2026 by researchers at Universitas Muhammadiyah Semarang explores how artificial intelligence is stepping into this gap. The authors, Muhammad Munsarif and Richmond Ampah-Mensah, gathered and analyzed thirty-two recent studies to see if deep learning—a powerful type of computer program that learns from data—can accurately judge seafood freshness without touching it. They looked at a wide range of technologies, from standard cameras that take pictures of fish eyes and gills to advanced sensors that detect invisible chemical signals. The researchers found that while computers are becoming remarkably good at spotting spoilage, the path to real-world use is not as simple as achieving a high score on a test. The true measure of success, they argue, is not just how well a model predicts freshness in a controlled lab, but whether it can reliably do so on a busy factory floor or in a grocery store, using data that truly reflects the biological reality of the food.

The review began by sifting through more than a thousand scientific records published between 2021 and 2026. After removing duplicates and filtering out studies that did not use deep learning or relied on destructive testing, the team focused on twenty-eight high-quality papers. These studies covered a diverse array of seafood, including finfish, shrimp, crabs, and oysters. The researchers identified four main ways these systems "see" freshness. The first and most accessible method uses standard red-green-blue cameras, similar to those on smartphones, to analyze visible changes. These systems look for shifts in color and texture on the fish's skin, eyes, or gills, which often turn cloudy or discolored as the animal spoils. The second approach involves spectroscopy, where sensors shine light on the fish and measure how the tissue absorbs or reflects different wavelengths. This allows the computer to detect molecular changes, such as the breakdown of proteins or fats, that happen long before the fish looks bad to the human eye.

The third category involves intelligent sensors that act like a digital nose or tongue. These devices use chemical strips that change color when they react with gases released by rotting fish, such as ammonia. A camera then reads these color changes and feeds the data into a computer program. The fourth group combines these methods, using both visual images and chemical signals to get a fuller picture of the fish's condition. Across these different technologies, the reported performance was often impressive. Many studies claimed classification accuracy rates above 95 percent, meaning the computer correctly identified whether a fish was fresh or spoiled in nearly every case. Some systems that predicted specific chemical values, such as the level of volatile nitrogen, showed a very strong statistical match to the actual laboratory measurements.

However, the authors caution that these high numbers tell only part of the story. A major finding of the review is that many studies rely on "labels" that are too easy to guess. For instance, if a computer is trained to distinguish between fish stored for one day versus fish stored for ten days, it might simply learn to recognize the lighting conditions or the background of the lab rather than the actual state of the fish. The review found that studies which anchored their predictions to hard chemical data, such as the count of living bacteria or the specific levels of spoilage metabolites, provided much stronger evidence of real-world utility. When a model predicts a specific chemical value rather than just a storage time, it is more likely to be measuring the actual biological process of decay.

Despite the promising accuracy, the researchers identified significant hurdles before these systems can be widely adopted. A critical weakness in most of the reviewed studies was a lack of independent testing. Many systems were tested only on the same batch of fish, in the same laboratory, under the same lighting conditions that were used to train them. This is like teaching a student to take a test by giving them the exact same questions they will see on the exam; they may get a perfect score, but they have not truly learned the subject. The review found that very few studies tested their models on fish from different batches, different species, or different locations. Without this kind of external validation, it is difficult to know if the system will work when the conditions change, such as when a fish is viewed under different store lights or by a different camera.

Another challenge lies in the difference between a computer model working in theory and a device working in practice. While many studies proposed that their systems could be put on a smartphone or a robotic arm, only a handful actually demonstrated this. A few researchers successfully ran their programs on portable hardware like Raspberry Pi computers or mobile phones, showing that the technology can be fast and small enough for real use. One study even demonstrated a robotic system sorting fish based on freshness. Yet, these practical demonstrations remain the exception rather than the rule. Most research is still stuck in the proof-of-concept stage, where the focus is on building a model that works in a controlled environment rather than one that can handle the messiness of a real supply chain.

The authors conclude that the field has moved beyond simple image recognition to a more complex integration of biology and technology. The most mature systems are those that pair the right sensing tool with a clear biological goal, such as measuring a specific spoilage compound, and then test that system rigorously outside the lab. High accuracy alone is not enough to declare a technology ready for the market. For these tools to truly transform how we handle seafood, future research must prioritize testing on diverse, independent samples and proving that the systems work on the actual hardware they will use in the real world. Until then, the promise of instant, non-destructive freshness checks remains a powerful potential rather than a solved problem, waiting for the next step from the laboratory bench to the fish market.