LCNet: Lightweight real-time image classification network based on efficient multipath dynamic attention mechanism and dynamic threshold convolution

Abstract

Hybrid architectures that integrate convolutional neural networks (CNNs) with Transformers can comprehensively extract both local and global image features, exhibiting impressive performance in image classification. However, their large parameter sizes and high computational demands hinder deployment on low-resource devices. To address this limitation, we propose a dual-branch classification network based on a pyramid architecture, termed LCNet. First, we introduce a dynamic threshold convolution module that adaptively adjusts convolutional parameters based on the input, thereby improving the efficiency of feature extraction. Second, we design a multi-path dynamic attention mechanism that optimizes attention weights to capture salient information and enhance the significance of key features. Third, a star-shaped connection is adopted to enable efficient information fusion between the two branches in a high-dimensional implicit feature space. LCNet is evaluated on four public datasets and one wood dataset (Tiny-ImageNet, Mini-ImageNet, CIFAR100, CIFAR10, and Micro-CT) using recognition accuracy and inference efficiency as metrics. The results show that LCNet achieves a maximum accuracy of 99.50% with an inference time of only 0.0072 s per image, outperforming other state-of-the-art (SOTA) models. Extensive experiments demonstrate that LCNet is more competitive than existing neural networks and can be effectively deployed on low-performance computing devices. This broadens the applicability of image classification techniques, aligns with the trend of edge computing, reduces reliance on cloud servers, and enhances both real-time processing and data privacy.