FPGA acceleration of automated ship detection and CNN-based sip/iceberg discriminator in SAR imagery

The need for automated ship detection methods has become increasingly important with the advancements in Synthetic Aperture Radar (SAR) technology in Maritime Domain Awareness in Canada. In this thesis, we present an automated ship detection algorithm for SAR imagery based on a Trimodal Discrete Model and Nelder-Mead Simplex Algorithm. We explain the theoretical foundation of the algorithm and its optimization techniques to improve its performance. Furthermore, we present the FPGA implementation of this system, which improves its speed and efficiency. Since ships and icebergs can appear similar in SAR images, we design and train a Convolutional Neural Network (CNN)-based classifier to discriminate between these two objects. To make the CNN model suitable for deployment on small devices, we apply network quantization to shrink its size. Our results demonstrate that the quantized model with 8-bit weights and activation functions has the same accuracy as the floating point one. Overall, this thesis provides a comprehensive solution for automated ship detection in SAR imagery, including a novel statistical model, FPGA implementation, and deep learning-based classification. Our approach improves the accuracy and efficiency of ship detection and classification in SAR imagery, which has practical implications for maritime surveillance and safety.