Infrared Image Classification and Detection Algorithm for Power Equipment Based on Improved YOLOv10

• Published in: IEEE access Vol. 12; p. 1
• Main Authors: Ji, Xiu, Yue, Zheyu, Yang, Hongliu, Zhang, Zehong
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Background noise; Classification algorithms; Computational modeling; Convolution; Convolutional neural networks; Entropy (Information theory); Fault diagnosis; Feature extraction; Feature recognition; Image classification; Improvement of YOLOv10; Industrial applications; Infrared imagery; Infrared imaging; Infrared Imaging of Power Equipment; Lightweight; MDCR; Modules; Neck; Real-time systems; Slim Neck-GSConv; Target detection; Weight reduction; YOLO
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3514103

Real-time target detection has important applications in industry. Identifying the type of power equipment in massive infrared image data for heat fault diagnosis has emerged as a crucial issue that needs to be addressed in this field. However, infrared imaging technology has shortcomings such as poor signal clarity and serious background noise interference. To address this problem, this paper proposes an infrared image classification and detection algorithm for power equipment based on the improved YOLOv10, named YOLOv10plus. Firstly, a lightweight convolution technique (GSConv) is implemented within the central network system to maximize the retention of target features. Secondly, a Slim Neck design structure is used in the neck network and combined with a dual convolution module (DualConv) to achieve a lightweight model. Finally, a Multi Cavity Channel Refiner (MDCR) module is implemented at the foundational level of the network to enhance the target feature recognition. To better handle the imbalance in sample distribution, the original Binary Cross-Entropy (BCE) loss function is modified to incorporate Focal Loss, aiming to enhance model performance by focusing more on hard-to-classify examples. The experimental findings indicate that the enhanced model achieves a substantial improvement in detection performance on power equipment infrared image data. Its FPS value reaches 588 and mAP value reaches 85.3%, which are 361 and 13.3% higher than the original model, respectively.