Performing Aedes mosquito eggs detection based on HSV segmentation and image classification. HSV segmentation and eggs localization was done on this repo and image classification for detection was carried out based on shape classification. Shape features were extracted using Elliptic Fourier Descriptor using pyEFD library. Common machine learning algorithms, such as Support Vector Machine, k-Nearest Neighbors, and Random Forest was trained using egg and not egg images. The 'not egg' class includes cropped egg and random shapes.
Since the dataset only has 30 images of each class, data augmentation was conducted. The augmentation was done by rotating each image 360 degrees and extracting features every 5 degrees. So each image will be augmented to new 72 images with angle variation.
Hyperparameters of each model were optimized using Gridsearch and cross-validation with hyperparameter settings and search space referring to a paper.
As the model was trained using the shape of an egg (which is oval) and random/cropped egg shape (not oval), there are often found overlapping eggs that were segmented as one object and detected as 'not egg' (since it's not oval-like in the image). To overcome this, the overlapping object separation method could be used to further analyze and detect each egg that is overlapped.
