This project was done as part of a Kaggle competition found here. I use deep learning combined with a home-made background removal technique to classify seedling images as one of 12 species, scoring ~90% accuracy.
This project marked the first time I’ve ever used a number of techniques including image augmentation and data generators which have proven very powerful in later work. Although I did not use a pre-trained models in the final iteration of this project, I gained experience playing with a few of these models and exploring the impact of trimming them off at various depths.
The most important lesson learned in this project was the value of an organized system for logging the details and results of experiments as I tuned hyper-parameters. Losing track of everything I tried was extremely frustrating and I am prepared to take a much better approach for my next project.
See full projects on my personal webpage.