DSF-Net: Dual-Space frequency dynamic network for spacecraft thermal control coating defect detection

Abstract

Thermal control coatings are critical functional materials for high-end equipment such as spacecraft, directly affecting their operational stability and service life. During manufacturing, process parameter fluctuations and material contamination can easily cause various defects including crystal points, blemishes, and wrinkles. Traditional detection methods struggle to meet industrial real-time and high-precision requirements. Existing deep learning algorithms for spacecraft thermal control coating detection face three major challenges: lack of standardized industrial datasets, complex industrial lighting conditions interfering with feature extraction, and decreased recognition accuracy when multiple defects coexist. To address these issues, this paper first establishes an industrial-grade thermal control coating defect dataset (ITCCD), covering six typical defect types including crystal points and blemishes. Through frequency domain analysis, this paper reveals the spectral feature distribution patterns of different defect types and establishes the relationship between defect characteristics and frequency domain representation. Based on these findings, this paper designs a Dual-Space Frequency Dynamic Network (DSF-Net) with two key technical innovations: (1) a Dynamic Inception Mixer backbone network using cascaded dual-stage architecture with adaptive weight allocation mechanism; (2) a frequency-aware feature fusion module that implements frequency domain mapping and decoupling through learnable frequency response functions. Experimental results on the ITCCD dataset show that DSF-Net achieves 83.0 % on the mAP50, outperforming the best comparison model by 7.4 %, while reducing parameters by 23.9 %. Notably, detection accuracy for challenging defect types such as crystal points and blemishes improved by 22.7 % and 13.3 % respectively. This research constructs a standardized dataset and improves defect detection accuracy under complex lighting conditions through a frequency domain analysis framework. The proposed frequency domain feature decoupling method provides a new approach for material surface defect detection. DSF-Net achieves end-to-end detection while maintaining high accuracy, providing an effective technical solution for online quality control of thermal control coatings.