Classifying Dogs vs Cats - Edge Extraction

Edge Extraction

Another popular type of feature extraction in images includes edge extraction. This is very applicable to our dataset because cats and dogs have very distinct shapes. For example, cats usually have small pointy ears while dogs have larger floppy ears or ears that are more rounded. Furthermore, cats also have much larger whiskers compared to dogs. Therefore, if we can feed edges as extracted features into our ML models we can save a lot of computation power and time, while increasing accuracy.

To perform edge extraction on our images, we used the python CV2 library to convert images into the grayscale and then used the Canny Edge Detection Algorithm. Once we identified the edges we enhanced them so that we would be able to better distinguish between cats and dogs.

Canny Edge Detection Algorithm

The Canny Edge Detection is a popular multi-step algorithm that is used to detect various types of edges in an image. Here is a brief overview of the various steps in the algorithm.

A Gaussian Filter Kernel is convolved with the image.

We used a 10x10 Kernel so that we would only be able to detect very large edges in our image. This is because most of our images have very noisy backgrounds and we didn't want our filter to pick up edges in those parts of the image.

Find the intensity gradient of the image.

In this step, the algorithm uses filters to find horizontal, vertical, and diagonal edges in the image.

Gradient Thresholding

This is used to find areas where there is the greatest change in intensity value. These areas are usually where the edges are located.

Double Thresholding

This step is used to eliminate extra noise and make sure we only detect the most significant edges.

Hysteresis Edge Tracking

In this step, we make edge connections using blob analysis.

Above are a few images going through the process of edge extraction.

Left: original picture after it has been resized, Midde: Gray-Scale picture, Right: Image after Edge Detection

PCA

Below are some plots after perfoming a 2-component and 3-component PCA.

Here is a link to our code:

PCA-Edge

Based on these graphs, one might initially think that the PCA tells us that cats and dogs can't be distinguished by edges. However, if one looks closely at the 2-D graph they can see V-shaped pattern (<). Therefore, we also performed the LDA.

LDA

Below we have performed the LDA using Edges as Features.

Here is the link to our code:

LDA-Edge

On the left, we have a plot after performing LDA on datasets of cats and dogs of size 300 each. As we can see, there is a clear difference between the cat and dog data sets. This indicates that edges are indeed a key feature separating the classes.

When using a larger dataset, we can really see the extremes in datasets. It may seem odd, that there are only 3 points. However, it is important to note that since we only have 2 classes(cats and dogs) LDA can only be performed on one dimension. Upon printing out the values, we noticed the values for cats were extremely close together, and the values for dogs were extremely close together as well. Recall, that in the PCA of 1000 images we see a V-shape of the data sets (<). In the LDA, the point of the V is what we are most likely seeing. In other words, think of the PCA as being projected into the screen.

That being said, when using the data sets of size 1200, there is a red dot in the middle. This dot is actually an overlap between a subset of cats and dogs, where our algorithm wasn't able to detect edges properly. These images were mostly black or white, and happened because the dog/cat took up the whole image, or because there was too much noise in the image.

Examples of these types of images are below:

In this image, the cat fur had the same pattern as the rug, so the algorithm wasn't able to detect it properly.

In this image, the cat was pretty much the entire picture. Since, our Kernel size was so big to eliminate noise it wasn't able to pick up the finer details within the cat. As a result, it is a black image.

This image, had a lot of noise even in the gray-scale version. As a result, our pre-determined Kernel size wasn't specific enough to accurately detect edges.

Altogether, it was very hard to determine a kernel size for edged detection for our dataset, because it included so many types of images. For example, images where the cat or dogs face is the entire picture would require a very small Kernel Size to detect edges. This is because we are contrasting several parts of the cats/dogs with itself. Meanwhile, images where the there is a lot of noise and the cat and dog is a small part of the image would require a very large Kernel size, because we do not want any of the background noise. Using a medium sized Kernel worked for our purposes, but missed out on edge cases.

In the future, it would be beneficial to use a model that can determine the kernel size based of each image when doing the edge extraction. Models like the VGG-16 would be very useful, but feature extraction using it is beyond the scope of this project due to our lack of computational power and memory.

Note: The maximum sizes of cat and dog data sets we could pass into to the PCA was 1000, and 1200 for the LDA before running out of RAM on Google Colab.

Page updated

Report abuse