CNN-Transfer_Learning

2 different versions of Image Classifier were used. Metric used to compare their performance was validation accuracy.

Transfer Learning : VGG16

The model can be imported from keras itself. Last FC layers were popped. New trainable FC layers are added. All previous layers were frozen. Validation accuracy : >90%.

CNN Image classifier

An ordinary CNN model was used. Model has more than 1 Conv layers. Validation Accuracy : 80%.

Conclusion

Transfer Learning proved to be more accurate with 90-100% accuracy. CNN models don't always have such high accuracy. Alot of time is wasted in choosing the hyper parameters. So many trials and errors. Whereas in the case of Transfer Learning, you always get such high-accuracy. This is why Transfer Learning is the most effective method of solving computer vision problems.