A lightweight cascaded CNN architecture for efficient image classification in edge platforms

Isuru Bandara Samarakoon

A lightweight cascaded CNN architecture for efficient image classification in edge platforms

Publication Date

July 25, 2026

Creator

Isuru Bandara Samarakoon

University of Nottingham, Malaysia

Abstract

Convolutional neural networks (CNNs) have demonstrated remarkable success in image classification, driving their widespread adoption in research and industry. While deeper networks introduce through more complex computations to achieve higher accuracy, they also incur longer inference times. This creates a fundamental trade-off between accuracy and inference latency, as increased model complexity slows down individual classifications and, consequently, the overall performance of vision-based systems. This thesis presents a novel cascaded CNN architecture designed to achieve an optimal balance between classification accuracy, inference speed, and computational efficiency for real-time operation on edge devices. The proposed two stage design integrates a lightweight main CNN for rapid initial classification and a secondary CNN that is selectively invoked to refine predictions from the main network. The architecture incorporates several optimization strategies, including efficient layer design, layer fusion, pruning, and hardware aware design, enabling the model to operate effectively within the limited memory and processing capabilities of embedded systems. The model was developed and trained using the MATLAB Deep Learning Toolbox, with performance evaluated across 200 object classes relevant to autonomous guidance and service robotics in the intended application of an airport environment. The proposed CNN model was trained on a dataset derived from the ImageNet ILSVRC-2012 dataset, using 70% of the data for training and 30% for validation. The test dataset comprised randomly selected images from the internet, as well as real-time images captured by a camera connected to the computer. The performance of the proposed cascaded neural network model was evaluated based on classification accuracy and execution time. Experimental assessments were conducted on both a GPU enabled workstation and a Raspberry Pi 4B to evaluate real-world performance. For a wide range of test data presented, the proposed network architecture enhanced the inference speed by 6.43% compared to the AlexNet while maintaining a 95% Top-1 classification accuracy for tests conducted on 200 object classes and proved to be within the expected performance criteria. Additional benchmarking across multiple state-of-the-art CNN models shows that the proposed system offers superior efficiency in terms of accuracy and inference, underscoring its suitability for real-time robotic guidance and other embedded vision applications. Furthermore, the proposed network model was tested on standard computer systems and Raspberry Pi models, demonstrating deployability and compatibility on edge devices where the test data suggest that the proposed fast executing network model is best suited for edge devices with a minimum computational resources environment.

Item Type

ethesis

Thesis Type

PhD

Supervisors

Shanmugam, Anandan

University of Nottingham, Malaysia

Al-Murib, Haider Abbas Mohammed

University of Nottingham, Malaysia

Keywords

CNNs

Cascaded CNN