Astro Sentai: Pattern Recognition

How do we classify plants? We usually look at the shapes of their leaves or observe the color of their fruits. As I recall from my previous Biology classes, the leaves vary from one plant to another, may it be the base, apex, or outline of the leaves. Also, we expect that the color of the fruit is the same as the color of the flower. Fruits with a dark red or violet to dark violet colors are rich in anthocyanin, which increases the amount of anti-oxidants in our body. So the next time you buy apples, choose only those with a dark red color. Those that have a red with yellow patches tend to be older and have lost their anthocyanin.

Anyway, I'm a physicist so I might be wrong :D What I'll be discussing in this blog is recognizing patterns from images. I captured photos of two types of flowers in UP Diliman:

Figure 1. Blue plumbago (Plumbago auriculata)

Figure 2. Galphimia vine (Galphimia gracilis)

I didn't know what these flowers are. I had to post it on my Facebook account and wait for someone to name them. Luckily, my friend Frances was familiar with these flowers! *phew* It saved me a lot of time looking for their names.

My goal is to have a measure on these petals in such a way that when I have one of them, I'll be able to tell what flower it is. But instead of picking the flowers and tearing the leaves apart, I just took photos of them at different angles and lighting.

Figure 3. Training set for the blue plumbago

Figure 4. Training set for the galphimia vine

No, I really can't pull out the petals even for the sake of the activity. I used GIMP to remove the petals from the flowers to have 10 training images for both flowers. It's difficult to use image segmentation based on the colors, as it will locate the entire flower from the image. Brute force was needed to have training sets.

I characterized the petals according to:

SHAPE - well not actually shape, as I found no means to measure the shape or determine the type of its base, apex, or what. I measured the ratio of the area of the blob (which is simply the number of nonzero pixels) to the area of the region enclosing the blob (from the AnalyzeBlob function of Scilab, I got the coordinates of the bounding box). Figure 5 below illustrates the regions of interest for both type of flowers. As you would see, the blue plumbago occupies more region on the bounding box than the galphimia vine because it is more round. The epithet auriculata pertains to the ear-like feature of the petal [1]. The petal of the galphimia vine has a pointed base, allowing more black pixels to occupy the bounding box.
COLOR - from the normalized chromaticity coordinates, I measured the distance of each data point (on the NCC) from the origin and summed them all up. I know this is a weird method, but then again I can't think of any other way to have a measure of the color.

Figure 5. Estimate of the shape of the two types of petals

Figure 6. Quantifying the values from the normalized
chromaticity coordinates

From these classifiers, I plotted the training sets:

Figure 7. Locating the training sets on the color versus shape plot. Blue dots indicate the training images
for the blue plumbago flowers, yellow dots indicate those of the galphimia vine flowers.

Now that I have a clustered plot (the training sets are quite grouped close to each other), I will now try to locate these three test petals:

Figure 8. Test images. The symbols below indicate the legend for the next plot.

Using the same measures for the shape and color:

Figure 9. Locating the test images.

Based on Figure 9, by visual inspection, we can say that the two blue plumbago test petals are closer to the majority of the blue dots. Similarly, the test petal for the galphimia vine petal was closer to the yellow dots. Using simple distance measurements:

Table 1. Distance of the test images from the training sets

You'll see that both the first and third test images (both blue plumbago petals) have smaller distance from the blue compared to the yellow, while the second image (galphimia vine petal) has a smaller distance from the yellow compared to the blue. This shows that using the two classification schemes, I was able to determine what flower our test images come from.

For this activity, I will give myself a grade of 12/10.

EDITED:
I tried another classifier for the shape: