@article{Fan2026, 
author = {Jia Fan and Peng Wang and Yue Li and Yutai Lu},
title = {Lightweight operation step recognition network based on post enhanced depthwise convolution and spatial-frequency attention},
year = {2026},
journal = {Journal of Measurement Science and Instrumentation},
volume = {17},
number = {3},
pages = {395-406},
keywords = {visual recognition, image classification, lightweight convolutional neural network (CNN), post enhanced depthwise (PED) convolution, spatial-frequency attention (SFA) mechanism, channel shuffle, difficult-to-classify classes},
url = {https://www.sciopen.com/article/10.62756/jmsi.1674-8042.2026034},
doi = {10.62756/jmsi.1674-8042.2026034},
abstract = {In practical operational training, image classification is employed to recognize trainees’ operational results in real time. Based on the recognition results, effective feedback can be provided to trainees, which plays a critical role in improving teaching quality and helping mitigate the shortage of teaching resources. Lightweight convolutional neural networks (CNNs) can satisfy real-time recognition requirements. However, recognition accuracy remains low for classes with highly similar operational steps. To address this issue, we propose a lightweight operation step recognition network based on post enhanced depthwise (PED) convolution and spatial-frequency attention mechanism. A post enhanced depthwise convolution module is introduced to effectively improve the nonlinear representation capability of the network, and a spatial-frequency attention (SFA) mechanism is designed to strengthen feature calibration ability, thereby effectively enhancing and integrating important channel features with spatial dimension information. The initial convolution layer and final convolution layer of the backbone network are jointly optimized, reducing the number of model parameters and improving feature information flow. Experiments were conducted on a self-built dataset with ten standard operational step classes. The results show that the proposed method increases the accuracy of two difficult-to-classify classes to 93.50% and 96.02% respectively, the overall Top-1 accuracy in ten classes reaches 95.69%, the training time is only 198 s, and the single-frame inference latency on edge devices is 4.93 ms. These results demonstrate an optimal balance between performance and model lightweightness, superior to other mainstream lightweight networks. The designed network has lightweight and efficient network architecture, and high sensitivity to subtle differences. Therefore, it is appropriate for the devices with limited computing power, and it also has the potential to generalize other real-time recognition tasks with clear steps and indistinct inter-class differences.}
}