Contextualizing CNN attribution maps using predefined image-derived descriptors in medicinal plant seed classification

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking at a picture. In the world of artificial intelligence, computers are getting really good at looking at photos and guessing what they are. This is called "image classification." If you show a computer a picture of a dog, it might say, "That's a dog!" with 99% confidence. But here's the tricky part: the computer doesn't actually "see" like we do. It's a bit like a magic black box that spits out an answer without telling you why it chose that answer. Did it look at the ears? The tail? Or maybe just the background?

To fix this, scientists use something called "attribution maps." Think of these as a high-tech highlighter pen. When the computer makes a guess, the attribution map lights up the specific parts of the image that the computer thought were most important. It's like the computer is pointing a finger and saying, "I saw this spot, and that's why I made my choice." But there's a catch: the highlighter just shows a blurry, glowing blob. It tells you where the computer looked, but not what it saw in that spot. Was it the color? The texture? The shape? That's where this new study comes in, trying to translate that blurry glow into a language we can actually understand.

---

The Mystery of the Medicinal Seeds

In this study, a team of researchers from Kyung Hee University decided to tackle a very specific, very tricky mystery: telling different types of medicinal plant seeds apart. These seeds are tiny, and some of them look almost identical to the naked eye—like twins who wear the same clothes. If a computer gets them wrong, it could be a big problem for people who rely on these seeds for medicine.

The researchers gathered a collection of 1,124 photos of five different kinds of seeds: Armeniaca vulgaris, Armeniaca sibirica, Prunus japonica, Prunus davidiana, and Prunus persica. They taught a super-smart computer brain (a type of AI called a Convolutional Neural Network, or CNN) to look at these photos and sort them out. The computer was incredibly good at it, getting the answer right almost every single time. In fact, three different computer brains they tested were so accurate they reached what the authors call "near-ceiling performance," meaning they were basically perfect at the task.

But the researchers didn't stop at just saying, "The computer is smart." They wanted to know: What is the computer actually looking at?

The "Highlighter" vs. The "Visual Dictionary"

To answer this, the team used a tool called Integrated Gradients (IG). Imagine the computer is a detective looking at a seed photo. The IG tool creates a "heat map" that glows red where the detective is staring hardest. But as we mentioned, a red glow doesn't tell you if the detective is staring at a shiny spot, a rough texture, or a specific color.

So, the researchers invented a clever way to decode that glow. They created a set of "predefined image-derived descriptor maps." Think of these as a visual dictionary or a recipe book for the image. They broke the seed photo down into simple, measurable ingredients:

• Color and Intensity: How bright is it? How light or dark? (Like measuring the "lightness" or "brightness").
• Spatial Frequency: Is the pattern smooth and blurry, or is it jagged and detailed? (Like measuring "coarse" vs. "fine" details).
• Texture: Is it bumpy or smooth?
• Edges and Shapes: Where are the outlines?

They then took the computer's "glowing highlighter" (the attribution map) and compared it to their "visual dictionary" (the descriptor maps). They asked a simple question: Does the computer's highlighter glow in the same places where the "brightness" recipe is strong? Or does it glow where the "texture" recipe is strong?

The Big Discovery: It's All About Light and Low-Frequency Details

After crunching the numbers, the researchers found a very clear pattern. The computer wasn't looking at the complex, jagged edges or the tiny, high-frequency details. Instead, the computer's "highlighter" was glowing brightest in the exact same spots where the lightness (how bright the seed is) and low-frequency details (the smooth, broad shapes) were strongest.

Specifically, the study found that the computer's attention was most strongly linked to:

1. LAB L (a measure of perceptual lightness).
2. Brightness (the overall intensity of the light).
3. FFT low-pass (which captures the smooth, coarse-scale variations in the image).

In simple terms, the computer was mostly looking at how light or dark the seed is and its general, smooth shape, rather than getting bogged down in tiny, noisy textures. The researchers also checked if other things, like specific colors (saturation) or complex outlines (Fourier descriptors), mattered. They found that the computer didn't seem to care much about those; in fact, for some of those features, the computer's attention was actually negatively linked, meaning it ignored them.

The "Summary" Test

To double-check their findings, the researchers did a second experiment. They took all those "visual dictionary" ingredients (the brightness, the texture, the edges) and turned them into a simple list of numbers (summary statistics). They fed this list into a different, simpler type of computer brain to see if it could still tell the seeds apart.

The result? Yes. Even without the fancy deep-learning model, just using the summary numbers from these visual ingredients allowed the computer to sort the seeds with very high accuracy (about 97.8%). This proved that the visual information the computer was using was actually there and was enough to solve the mystery on its own.

Why This Matters

This study is like giving the computer a voice. Before, we just knew the computer was right. Now, we know how it's right. By showing that the computer relies on lightness and broad shapes, the researchers have created a new way to check if an AI is "thinking" correctly. If a computer starts highlighting the wrong things (like the background or a random speck of dust), we'll know it is not relying on the intended features.

The authors suggest that this method—comparing the computer's "glow" to a "visual dictionary"—is a powerful tool for making sure AI is trustworthy, especially when dealing with tiny, tricky things like medicinal seeds where getting the answer right is crucial. They didn't just find a way to sort seeds; they found a way to peek inside the computer's brain and see what it's really seeing.