• machine learning
• keypoint detection and description
• computer vision
• convolutional neural networks
• deep features

This Thesis investigates using convolutional neural networks as a keypoint detector and descriptor for Computer Vision tasks, and proposes an original identification scheme named KIDD. To realize the system, State of the Art solutions have been examined, a framework has been designed and developed, three different test datasets were created and different models were trained on each, all able to predict the existence of a keypoint in a certain patch. Non-maximal suppression was applied in order to retrieve a balanced number of keypoints, and deep features were extracted and used as descriptors for keypoint matching between similar images. All models have been evaluated on every other dataset, and tests regarding the number of found keypoints and their match have been analyzed for a certain number of image pairs and compared to existing STAR methods.