Analysis of Hardware Accelerated Deep Learning and the Effects of Degradation on Performance

As convolutional neural networks become more prevalent in research and real world applications, the need for them to be faster and more robust will be a constant battle. This thesis investigates the effect of degradation being introduced to an image prior to object recognition with a convolutional neural network. As well as experimenting with methods to reduce the degradation and improve performance. Gaussian smoothing and additive Gaussian noise are both analyzed degradation models within this thesis and are reduced with Gaussian and Butterworth masks using unsharp masking and smoothing, respectively. The results show that each degradation is disruptive to the performance of YOLOv3, with Gaussian smoothing producing a mean average precision of less than 20% and Gaussian noise producing a mean average precision as low as 0%. Reduction methods applied to the data give results of 1%-21% mean average precision increase over the baseline, varying based on the degradation model. These methods are also applied to an 8-bit quantized implementation of YOLOv3, which is intended to run on a Xilinx ZCU104 FPGA, which showed to be as robust as the oating point network, with results within 2% mean average precision of the oating point network. With the ZCU104 being able to process images of 416x416 at 25 frames per second which is comparable to a NVIDIA 2080 RTX, FPGAs are a viable solution to computing object detection on the edge. In conclusion, this thesis shows that degradation causes performance of a convolutional neural network (quantized and oating point) to lose accuracy to a level that the network is unable to accurately predict objects. However, the degradation can be reduced, and in most cases can elevate the performance of the network by using computer vision techniques to reduce the noise within the image.