When we should use the pre-trained model as deep features?



Nowadays deep learning used in the most of the computer vision applications and it improves predictions of the system. In the old days, we have spent most of the time on feature selection (Handcrafted) and doing the classification or segmentation with these features. We could use the pre-trained networks as a features extractor to do the classification and object detection.

Based on the samples count , we are using the pre-trained networks as a feature extractor.

  • If the training samples are less than 1000, you should use the deep feature for classification. (i.e, document images, histopathology images )
  • If you are working on image retrieval , you can use the deep feature. No need to spend time on training the last fully connected layer.
  • Deep features were used in face recognition and image clustering applications.

Source: Deep Learning on Medium