Birdometer

Overview

The plan is to test out different classifiers algorithms to detect if there is a bird in the image or not. Currently implemented, as a baseline, is a kNN classifier using KNeighborsClassifier from sklearn. Next up are logistic regression, neural network and support vector machine. Then, once the best classifier has been identified , I will train and use the test-set to understand the precision of the classifier. It will then be deployed on the Raspberry Pi to classify new images while in memory. The idea is to not save any image that does not have a bird. Depending on the performance, a continuous image-capture mode can be tested an implemented.

The current code can be found in the Jupyter Notebook here.

A closer look at the data
By doing dimensionality reduction, we can try to understand the data a bit better. Is there a clear separation between bird/non-bird pictures? If not, why not? What are the problem images and is there any other processing we can do to improve separation?

As a starter, a principal component analysis (PCA) uses a Singular Value Decomposition of the data to project it, in our case, to a two dimensional space. sklearn.decomposition.PCA is used to find the principal components and the first two principal components are plotted. The following snippet is adapted from the sklearn example. X_train and y_train are the features and labels (0 or 1) for the training set.

from sklearn.decomposition import PCA
pca = PCA(n_components=2, svd_solver='randomized',
          whiten=True).fit(X_train)
birds_pca = pca.transform(X_train)
# birds_pca is a NumPy array that contains the two components for each image.
# [[-0.91417182 -0.58942785], ..]
# The rest is basically just for plotting
colors = ["#27AE60", "#E74C3C"]
plt.figure(figsize=(15,5))
# Set the axes limits, accounting for the marker sizes as well
plt.xlim(birds_pca[:,0].min(), birds_pca[:,0].max() + 0.05)
plt.ylim(birds_pca[:,1].min(), birds_pca[:,1].max() + 0.2)
for i in range(len(birds_pca)):
    plt.text(birds_pca[i,0], birds_pca[i,1], str(y_train[i]), color=colors[y_train[i]])
plt.xlabel("pca1")
plt.ylabel("pca2")

Which gives the following result:

Plot showing the two components (pca1 and pca2) as extracted by the PCA algorithm. In green (red) are the images without (with) a hummingbird.

From the two-dimensional plot showing the two principal components above, it is clear that there are different clusters of images and within each cluster, both class 0 (no bird, green) and class 1 (bird, red) are represented typically. We can understand this as follows:

• The images have differences in shading and contrast depending on whether the sun is out or on the time of day.

• Each cluster is likely a cluster of images taken in similar lighting (time of day, shade, sun).

• Within a group of images taken in similar conditions, there are images with and without birds.

Next, it makes sense to look at a few images to investigate. Specifically, there are two interesting groups:

• group 1: pca1 < -0.5 and pca2 > 0.5. The classes are pretty well separated in this grouping.

• group 2: pca1 > 1.0 and pca2 < 0.0. The classes seem less well separated in this grouping.

This following plot shows group 1 in blue and group 2 in purple.

Plot showing the two components (pca1 and pca2) as extracted by the PCA algorithm. In green (red) are the images without (with) a hummingbird.

When looking through examples of both classes in this groups, it is clear that the shade and background play an important role in how well the classes can be separated. See examples below.

Group 1 class 1 (left) and class 0 (right) typical examples.

Group 2 class 1 (left) and class 0 (right) typical examples.

For group 1, there is a large contrast between the bird (in the shade) and the background (sunny). The distance in the (pca1, pca2) space is therefore larger than for group 2, where the contrast is minimal. A change to the hardware, assuring that the bird is always in the shade, can help in this example.

The next step is to implement different classifiers.

[This project page will be updated with more insights and results as they are being uncovered and produced.]