DSConvNet: A Lightweight Architecture for Extracting Image Features From Depthwise Separable Convolution Network for Edge Devices

초록

This paper presents DSConvNet, a novel architecture based on depthwise separable convolutional blocks for efficient multi-class image classification. Despite progress in compact convolutional neural networks (CNNs), many existing models still impose high computational costs on resource-constrained devices, limiting their real-time applicability. To address this, DSConvNet employs four optimized DSConvNet blocks that reduce parameters, accelerate inference, and stabilize training. Each block combines a 2D convolution, depthwise convolution, batch normalization, and max-pooling, forming an efficient yet expressive feature extractor. The resulting architecture achieves lower complexity, mitigates overfitting, and improves generalization. DSConvNet was evaluated on nine benchmark datasets GTSRB, BTSC, CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, Imagewoof, Imagenette, and Caltech-101 covering 100 object categories. Without GPU acceleration, the model attained 99.10% accuracy on GTSRB and 98.90% on BTSC, confirming its suitability for real-time edge-based image classification under limited computational resources.

키워드