Automatic Generation of Custom Image Recognition Models

Abstract

In this thesis, we examine the viability of training a convolutional neural network using synthetic data. The cnn is used for image recognition in an rms. We create a program that can render 3D models from STL files as images with varied backgrounds. The process of creating and training an image recognition model is also automated. Lastly, the model is used for image recognition. The report compares different methods and hyperparameters used in training a model. Transfer learning is found to be suited for synthetic datasets. Using the pre-trained feature extraction layers of the VGG-16 model, we train an image recognition model with better than 90% accuracy in the laboratory. We then demonstrate the use of this model for object detection, and suggest avenues for further development.