Improving CNN performance
One of the main objectives of the AI Trees project is to attain the optimal CNN performance. I trained instances of ResNet 50 and Inception V3 which allows us to compare SafeTree with state-of- the-art models. These new models, however, require fine-tuning in order to achieve better results. The preliminary results point at approximately 67% validation accuracy. It is important to point out the “Pr_Po_Im” scenario was used for training which typically leads to worse results given the three-class problem. Additionally, I generated performance summary metrics and classification reports for the different archutectures.
After training ResNet 50 and Inception V3 on multiple resolution types and scenarios, I discovered that deep networks tend to require much longer training (more than 100 epochs) for convergence to occur. Moreover, a three-class problem such as the Probable/Possible/Improbable classification still represents a challenge even for a state-of the-art model. Partially, this could be attributed to the size of the input data set as both Resnet and Inception typically require thousands of images for sufficient training.
By visualizing the CNN performance I observed that training reaches its limit between the 60th and 100th epoch. Early Stopping is a way to arrive at the optimal epoch values earlier and also avoid overtraining the networks.