Deep Learning in State-of-the-Art Image Classification Exceeding 99 Accuracy

摘要

Automatic image recognition and classification is a field of science that became popular in the recent years. Free platforms as Google Collaboratory make machine learning experiments more available to perform for everyone. The current technology enables us to use image recognition in such domains as medicine, criminology, entertainment or trading. In our research we created a state-of-the-art image classifier based on convolutional neural network model to classify ten models of old polish cars. We elaborated eight step training to fine tune the neural network. As the first step the data augmentation and precomputed activations were enabled. After that we froze all the layers but the last one, found the proper learning rate and performed interactive training. Fine-tuning, proper training and appropriate data preparation brought great results. The accuracy of cars' model recognition exceeded 99% with some room for improvement.