A Novel Lightweight CNN for Constrained IoT Devices: Achieving High Accuracy With Parameter Efficiency on the MSTAR Dataset

Convolutional neural networks (CNNs) have revolutionized fields such as image classification, natural language processing (NLP), and object detection. Remote sensing is no exception, particularly synthetic aperture radar (SAR) image analysis, where it addresses the challenge of classifying inherently noisy and indistinct images. Traditionally, this task relied heavily on manual intervention, making it time-intensive. With automation becoming increasingly critical, research has shifted toward deep learning techniques. However, many existing approaches are computationally intensive and require substantial memory resources. In this study, we present a lightweight CNN specifically designed for SAR applications, tested on the moving and stationary target acquisition and recognition (MSTAR) dataset. Our model surpasses previous studies by demonstrating higher accuracy while utilizing significantly fewer parameters. This novel architecture achieves an optimal balance between accuracy and computational efficiency. This is particularly useful in a resource-constrained environment in many real-world applications. More specifically, our proposed CNN model demonstrates robust performance across various scenarios, achieving an accuracy of 99.7% in classifying the three target classes under standard operating conditions (SOCs). Furthermore, when extended to classify ten classes, our proposed model outperforms several baseline algorithms from state-of-the-art literature. This research attempts to achieve a trade-off between model performance and model size, contributing to the development of CNNs suitable for resource-limited applications, particularly targeting IoT deployments. The findings present a practical solution for situations that prioritize both accuracy and resource conservation, thereby advancing discussions on efficient model design.