Fourier-Guided Multi-Scale Vision Transformer for Object-Independent Detection and Severity Grading of Rain-Induced Cracking in Sweet Cherry

Sweet cherries are a delicacy prized for their deep color, firm texture, and fleeting season, but their commercial success hangs on a fragile balance. A single heavy rainstorm near harvest can turn a premium crop into a loss. When water sits on the fruit's skin, it can force microscopic tears in the protective outer layer, creating cracks that may look harmless at first but quickly invite rot and water loss. For decades, farmers and packers have relied on human eyes to sort these fruits, a task that is slow, subjective, and often misses the tiniest fractures that matter most. These invisible flaws are the difference between a fruit that travels well to a distant market and one that must be discarded or sold for processing. The challenge is not just seeing the damage, but understanding that a crack is not a single thing; it is a progression from a hairline fracture to a deep split, and its location on the fruit changes how dangerous it is.

To solve this, researchers have turned to a new kind of computer vision that mimics how the human brain might look at a problem from two different angles at once. Instead of just training a computer to recognize patterns in an image like a standard photograph, this new approach teaches the machine to also "listen" to the mathematical frequencies hidden within that picture. In the world of images, smooth areas like a shiny cherry skin have low-frequency patterns, while sharp edges, fine lines, and tiny cracks create high-frequency signals. By combining a standard visual analysis with this frequency-based hearing, scientists have built a system that can spot the faintest signs of trouble before they become visible to the naked eye. This method was recently tested on sweet cherries to see if it could reliably grade the severity of rain damage and map exactly where the cracks are, offering a potential future where every single fruit is inspected with perfect consistency.

The study, conducted by a team of researchers, focused on creating a tool that could handle the messy reality of real-world fruit. They gathered 960 cherries from four different commercial varieties grown in various orchards. To ensure the system was tested fairly, they did not just take random pictures; they captured four different views of every single fruit, rotating them to show all sides. Some of these cherries had naturally cracked after rain, while others were carefully soaked in water in a controlled setting to create a wide range of damage, from tiny, almost invisible splits to deep, open wounds. Each fruit was then examined under a powerful microscope to measure the exact length and width of every crack, creating a strict "gold standard" of truth against which the computer's guesses could be measured. The researchers defined four levels of damage: intact, microcrack, moderate, and severe, based on precise measurements of how much the skin was broken and how much of the fruit's surface was affected.

The computer model they developed, which they call a Fourier-guided multi-scale Vision Transformer, works by looking at the fruit in two distinct ways simultaneously. First, it analyzes the image normally, breaking it down into small patches to understand the shape of the fruit, the position of the stem, and the overall color. This is the "spatial" view, which tells the computer where things are. Second, it takes the image and converts it into a frequency map, a mathematical representation that highlights changes in brightness and texture. This second view is particularly good at spotting the sharp, jagged lines of a crack, even if they are very faint or hidden by glare. The model then uses a special attention mechanism to let the frequency information guide the spatial view. It is as if the computer has a second pair of eyes that can see the invisible vibrations of a crack, and it uses that information to tell the first pair of eyes exactly where to look more closely. This allows the system to focus on the tiny details that usually get lost when a computer tries to simplify an image.

The results of this dual-approach were striking. The system correctly identified the severity of the damage on 93.8% of the fruits, a significant improvement over models that only looked at the image visually. More importantly, it excelled at finding the microcracks, the tiny fractures that are hardest for humans to see but most dangerous for the fruit's future. When the researchers tested the system's ability to draw a map of exactly where the cracks were, it achieved a high level of precision, successfully outlining the damaged areas even when they were less than a fraction of a percent of the total fruit surface. The model was also robust; it continued to perform well even when the fruit was partially hidden by its stem, when bright lights created confusing reflections, or when the image was slightly blurry. In fact, under difficult lighting conditions, the new model maintained a clear advantage over older methods, proving that the frequency-based "hearing" helped it see through visual noise.

One of the most valuable aspects of this work is that the computer does not just give a final grade; it also produces a visual map of the crack. This means a human operator can look at the computer's decision and see exactly which part of the fruit led to that conclusion. If the system flags a fruit as damaged, the operator can see the highlighted crack on the screen, making the process transparent and trustworthy. The study found that the computer rarely made wild mistakes; when it was wrong, it usually confused a "moderate" crack with a "severe" one, or a "microcrack" with a "moderate" one, rather than calling a healthy fruit broken or a broken fruit healthy. This suggests the system understands the continuum of damage rather than just guessing categories. The researchers also noted that the system learned to ignore irrelevant features like the color of the fruit or the background, focusing strictly on the structural damage, which is essential for a tool that must work across different varieties of cherries.

While the system shows great promise, the researchers are careful to note its current limits. The tests were conducted in a controlled environment with specific cameras and lighting, and the fruit were examined under ideal conditions. Real-world packing lines move much faster, and the fruit may be wet, dirty, or moving in ways that are harder to capture. The study also used a mix of naturally cracked fruit and fruit that was artificially soaked, and while this provided a wide variety of damage, the two types might not look exactly the same in every detail. Furthermore, the system was trained on four specific varieties of cherries, and it remains to be seen how well it will perform on new, unseen types of fruit. The researchers emphasize that before this technology can be used in a commercial packing house, it needs to be tested on a much larger scale with different equipment and under the chaotic conditions of a busy factory floor.

Despite these hurdles, the study offers a clear path forward for the future of fruit inspection. By combining the ability to see the big picture with the ability to hear the fine details, this new approach provides a way to sort cherries that is both fast and incredibly precise. It moves beyond simple counting or color sorting to a deep understanding of the fruit's physical integrity. For growers, this could mean less waste and higher quality fruit reaching the shelves. For scientists, it offers a new way to study how different cherry varieties react to rain, helping them breed tougher crops for the future. The work demonstrates that when we teach machines to look at the world in more than one way, we can uncover truths that were previously invisible, turning a simple image of a cherry into a detailed story of its health and resilience.