Design and performance analysis of an image classification model based on lightweight convolutional neural networks

Abstract

To address the issues of excessive parameter scale and sluggish inference speeds when deploying image classification models on resource-constrained devices, this study proposes a lightweight convolutional neural network (CNN) architecture and systematically investigates its structural optimization and performance enhancement. First, a comparative evaluation is conducted between Stochastic Gradient Descent (SGD) and Adam optimizers to analyze their impacts on convergence speed and classification robustness. Subsequently, a Fully Connected (FC) network, a two-layer convolutional network (CNN2), a two-layer convolutional network with doubled depth (CNN2W), a three-layer convolutional network (CNN3) and a four-layer convolutional network (CNN4) are developed to quantitatively analyze the trade-offs among parameter count, inference latency, and top-1 accuracy across varying network depths. On this basis, depthwise separable convolutions and Squeeze-and-Excitation (SE) attention mechanisms are integrated into the optimal baseline network (CNN2) to validate the effectiveness of lightweight structural variants. Experimental results demonstrate that the Adam optimizer equipped with a learning rate decay scheme yields superior training stability. Compared with CNN4, the shallower CNN2 architecture drastically reduces the model scale and inference latency while preserving high classification accuracy. Furthermore, the integration of depthwise separable convolutions significantly accelerates inference efficiency, achieving a favorable balance between inference efficiency, model complexity, and classification accuracy, whereas the SE attention mechanism exhibits marginal performance within the constructed small-scale network due to limited channel capacity. Ultimately, the finalized model, designated as CNN2D, achieves a refined balance among classification accuracy, computational footprints, and real-time inference efficiency. This work provides a practical and viable method for deploying low-complexity image classification networks on resource-constrained edge computing platforms.