LightClassNet-A Lightweight Convolutional Model for Efficient Image Classification on Edge Devices

Edge devices require more and more efficient models of deep learning algorithms that are able to deliver high levels of accuracy while following strict constraints in terms of computation and memory. Existing convolutional neural networks tend to compromise between performance and efficiency and there is a huge gap for real-time vision applications on resource-limited platforms. This study proposes LightClassNet which is a lightweight convolutional architecture to optimise the accuracy, latency and energy cost of edge-based image classification. The model is both strategy-wise by combining depthwise separable convolutions and bottleneck modules as well as pruning & quantization. Slightly improved results were observed in the experiments performed on the CIFAR-10 dataset indicating a 93.4% accuracy using only 2.1MB model size and 18ms inference time on a Raspberry Pi 4, outperforming several state-of-the-art lightweight models. The results confirm the proposed design as a good solution to fill the efficiency-accuracy gap, and an excellent solution to real-time, on-device AI applications.