Classifying Images with Transfer Learning

The sample TensorFlow iOS apps, simple and camera, and the Android app TF Classify, described in the previous chapter all use the Inception v1 model, a pretrained image classification deep neural network model made publicly available by Google. The model is trained for ImageNet (http://image-net.org), one of the largest and best-known image databases with over 10 million images annotated for object classes. The Inception model can be used to classify an image into one of the 1,000 classes, listed at http://image-net.org/challenges/LSVRC/2014/browse-synsets. Those 1,000 object classes include quite a few dog breeds, among many kinds of objects. But the accuracy for recognizing dog breeds is not that high, around 70%, because the model is trained for recognizing a large number of objects, instead of a specific set of objects such as dog...