An Instrument Indication Acquisition Algorithm Based on Lightweight Deep Convolutional Neural Network and Hybrid Attention Fine-Grained Features
Jiaquan Shen,Ningzhong Liu,Han Sun,Deguang Li,Yongxin Zhang
DOI: https://doi.org/10.1109/tim.2023.3346488
IF: 5.6
2024-02-02
IEEE Transactions on Instrumentation and Measurement
Abstract:Instruments detect and respond to various parameters, playing a crucial role in civil and industrial applications. However, the current instrument recognition algorithm still suffers from low detection accuracy, a large amount of model calculation, and a low degree of automatic recognition. In this article, we propose an efficient real-time automatic instrument indication acquisition algorithm based on a lightweight deep convolutional neural network and hybrid attention fine-grained features. In this algorithm, we divide instrument identification into three subtasks, which are instrument object detection, instrument fine-grained recognition, and instrument indication acquisition. In the instrument object detection task, we propose a lightweight instrument detection algorithm, which uses a lightweight feature extraction backbone network and a single-scale object detection algorithm to reduce the computational load. In the instrument fine-grained recognition task, this article proposes a fine-grained classification network based on attention hybrid clipping, which can effectively recognize the instrument fine-grained category and range features. In the instrument indication acquisition task, the edge detection and polar coordinate change are used to realize an automated instrument identification and reading system, which can meet the needs of high-precision real-time monitoring of the instrument indication in various complex scenarios. In addition, we have created and made publicly available an instrument image dataset that contains 3457 images of various types of instruments in actual scenes. The recognition accuracy of the instrument reading recognition algorithm proposed in this article is 92.9% on the created dataset, and the average recognition time per image is 348 ms on the i7-8700 CPU.
engineering, electrical & electronic,instruments & instrumentation