Aim: Is to apply ResNet-50 to classify different categories of flowers
Methods: This data set of the University Oxford (Visual Geometry Group - University of Oxford) contains n = 1360 flowers of 17 categories of various images sizes. We apply the ResNet-50 network architecture and evaluate accuracy with and without using a rejection rate. Training data set contains 1088 flowers and the test set 271 flowers.
Results: An illustration of the performance of this network architecture is given here:
The accuracy without a rejection threshold is 91.5% (+/- 1.7%) and increases to 96.6% (+/-1.2%) with a rejection threshold of 14.4%.
Discussion: This network architecture achieves a high accuracy, other architectures like Inception V3, Squeeze-and-Exitation, VGG 16 may even increase performance. Training was done on a notebook without a GPU and is fast for ResNet-50. More complex network architectures with larger amount of parameters may perform better, in this case we suggest to use a GPU. A MATHEMATICA 12 notebook with more details is attached.