Structural Prior Knowledge Integrated Deep Learning Method for Tiny Object Identification

With the rapid development of energy, manufacturing, and other industries, there has been a significant increase in the demand for equipment monitoring and maintenance. Taking an optical modem as an example, the connection status of the device can be monitored by analyzing its indicator labels and lights. However, detecting these indicators poses challenges due to their small size and potential interference. To achieve automatic equipment monitoring and enhance equipment maintenance efficiency, we propose a cascaded knowledge-integrated object detection method that leverages the inherent structural characteristics of the equipment. We focus on the detection region within the entire image and identify the object in this specific area. This approach effectively mitigates noise interference, enhancing both efficiency and accuracy in identifying regions of interest. Our method demonstrates strong scalability, making it suitable for status detection across various communication equipment. We evaluated our proposed method using datasets collected from real-world scenarios, ultimately achieving an impressive accuracy rate of 97.7%.