Simple_Classifier

Solution

Tried to build CNN from scratch but the extracting feature from a small dataset with random noise does not produce promising results.

• Clean the dataset by removing noisy images and adding few images for more support in a case where the model was performing worse
• Data Augmentation: Created more set of images by flipping, rotating, flipping and adding Gaussian noise to the image to create more data for the CNN to learn

• Use the pre-trained model trained on Imagenet dataset to fine-tune on the above dataset to produce a better result.

• to further increase the accuracy on the test set, use the average ensemble method consisting of InceptionV3, InceptionResnetV2 and Xception

Improvement:

• Clean the data more to remove the noise

• build an OCR in combination with Image classifier to cater to Food Categories containing words

• Results can be reproduced using the Github Repo

Built With

• fast.ai
• keras
• python
• tensorflow