The project aims to identify physical traits in plants using images as input and identify if traits within plant species are being affected by external factors.
The approach to analyzing the dependency of external factors on plant traits is partially supervised and partially unsupervised. The backbone of this algorithm is a Resnet50V2 model fine-tuned on the plant image data. The input plant image is segmented using Facebook’s segment anything model, then the segmented images are used to train the Resnet with additional layers using Transfer learning. Finally, the model is used to generate latent vectors of size 2048, decomposed to 100 using PCA and clustered using K means clustering. These individual clusters are then compared to provide analysis.
