This project is to built an image classifier for plant species by fine-tuning the top layers of a pre-trained CNN (Xception). To further improve the performance, a filter is created with OpenCV to mask the background of the images and improved the accuracy by 3%. Because the limit of the training data, an overfitting problem is observed and then solved by data augmentation and achieved 96% accuracy.

A training set and a test set of images of plant seedlings at various stages of grown. Each image has a filename that is its unique id. The dataset comprises 12 plant species. The list of species is as follows:

Black-grass Charlock Cleavers Common Chickweed Common wheat Fat Hen Loose Silky-bent Maize Scentless Mayweed Shepherds Purse Small-flowered Cranesbill Sugar beet

The dataset is available here https://vision.eng.au.dk/plant-seedlings-dataset/ and https://www.kaggle.com/c/plant-seedlings-classification/data

• OpenCV

• Plant_Seedlings_EDA.ipynb is the EDA of the dataset.
• ps_load_data.py loads the input data and generate pandas DataFrames contains the file paths, categories ids, categories, etc.
• ps_image_to_array.py process the training dataset without filtering the background.
• ps_image_to_array_filter.py process the training dataset and filter the background.
• ps_image_to_array_filter_test.py process the image data for the test dataset.
• ps_fine_tune_part1.py use the bottleneck features and train the dense layer.
• ps_fine_tune_part2.py re-train the top layers of the Xception model.
• main.py is the main function.

• Consider other pre-trained model, e.g. InceptionResNetV2.
• Generate more training data.
• Increase the number of layers to be re-trained.

Thanks to the authors below who provide excellent kernels:

https://www.kaggle.com/gaborvecsei/plant-seedlings-fun-with-computer-vision

https://www.kaggle.com/gaborfodor/seedlings-pretrained-keras-models