Date of Graduation

5-2026

Document Type

Thesis

Degree Name

Bachelor of Science in Computer Science

Degree Level

Undergraduate

Department

Computer Science and Computer Engineering

Advisor/Mentor

Dr. John Gauch

Committee Member

Dr. Lora Streeter

Second Committee Member

Dr. Qinghua Li

Abstract

Edge-optimized computer vision is a constantly evolving field where the definition of efficiency has changed repeatedly. This thesis presents a literature survey of four recent Convolutional Neural Network (CNN) families, all analyzed through a consistent framework of accuracy, parameter count, and Multiply-Accumulate Operations (MACs), alongside a survey of five CNN and Vision Transformer (ViT) hybrid models to examine the direction of the field. It was found that accuracy follows a logarithmic curve with respect to parameter count, exhibiting diminishing returns as models scale. This suggests that architectural design contributes more to performance gains than parameter count alone. Theoretical efficiency metrics such as Floating Point Operations (FLOPs) and MACs are poor indicators of real-world performance due to hardware-specific bottlenecks. Neural Architecture Search (NAS) consistently outperforms hand-designed models when the search space and target are well defined. CNN-ViT hybrid models are early in their developmental curve and are experiencing issues very similar to those of early CNN design. The boundary between CNNs and ViTs is dissolving, and future edge models will treat both as interchangeable tools rather than competing paradigms.

Keywords

Convolutional Neural Networks; Edge Deployment; Neural Architecture Search; Model Efficiency: CNN-ViT Hybrid Models; Hardware-Aware Design

Download

Share

COinS