86lekwenshiung / Multiclass-Image-Classification-with-CNN-in-Tensorflow

Using CNN in tensorflow and pretrained model from tensorflow hub (such as efficientnet and mobilenet) to train model to identify multiclass images.

Image Classification with Tensorflow CNN

1.0 Project Summary

---

Project	Findings	Key Feature	Data Source
Dog Breed Multi-classification	Summary	* Tensorflow * Transfer Learning * Convolutional Neural Network(CNN) * mobilenet_V2	Kaggle Dataset for Dog Breeds
Food 101 Multi-classification	Summary	* Tensorflow * Transfer Learning * Convolutional Neural Network(CNN) * Efficientnet_B0	101 Food Images

2.0 Findings for Dog Breed Multi-classification

---

Using pre-trained model from mobilenet_v2 (tensorflow hub) to train our model to identify ~ 120 dog breeds. Using a small data set of 1000 images to pre-trained our model , the model is observed to be very overfitted with a validation accuracy of 60% compared to a training accuracy of nearly 100%.

• Training Data : 800,
• Validation Data : 200

Using the same sequence , we increases the datasize to the full dataset of 10,000 images. The model performed much better , but it is observed to be overfitted with a validation accuracy of 81% compared to the training accuracy of 95%.

• Training Data : 8177,
• Validation Data : 2045
• Next Step : Further fine tuning layers could be added in the image preprocessing such as data augmentation and fine tuning the trainable layers.

Below image is a way we could visualize how well our model predict against the images ; The images where they performed well and not so well.