Date of Award

Spring 2023

Document Type

Thesis - SCU Access Only


Santa Clara : Santa Clara University, 2023.


Electrical and Computer Engineering

First Advisor

Hoeseok Yang


Satellite imagery plays an important role in many industries as it can be used to garner insights into countless trends. With the large amount of image data generated, modern object detection convolutional neural network (CNN) solutions must be utilized to label and apply bounding boxes around objects of interest. These systems must have high throughputs without sacrificing accuracy. Through our research, we analyze a state-of-the-art convolutional neural network (CNN) for satellite image object detection. We address limitations in hyperparameter tuning and mitigate a potential throughput bottleneck. As a result, we achieve impressive reductions in the model's inference time, of 39.23%, coupled with notable improvements in mean average precision (mAP) of up to 2%. Our work confirms the significance of random filter-based pruning in limited data CNN models. It mitigates overfitting, recovering 2.5% mAP and reducing weights by 73.34% with a minor inference time increase. Additionally we found that favoring less region proposals (generating only 250 proposals instead of originally 1000) with fewer training warm up cycles (only 300 cycles instead of 1000) resulted in a 10.30% decrease in inference time with a loss of mAP of only 2.43%. Finally our work identifies the non maximum suppression (NMS) portion of large object detection models as a major throughput bottleneck while exploring the performance of the modern Soft-NMS algorithm. While the Soft-NMS algorithm may not be the panacea for all object detection throughput optimizations as it can reduce or increase average precision depending on the image’s characteristics, its threaded implementation can better utilize most modern hardware to result in inference time reductions of up to 36% with negligible loss in accuracy. We propose optimizations and modifications to NMS solutions to further increase accuracy while retaining a high throughput.

SCU Access Only

To access this paper, please log into or create an account in Scholar Commons using your email address.