Radar-absorbing materials damage detection through microwave images using one-stage object detectors

An improved rapid multiprobe scattering microwave imaging algorithm and a deep learning model via microwave images are proposed to detect radar-absorbing materials (RAMs) damage in real time. The imaging algorithm improves the quality of microwave images by avoiding the transmitting and receiving antennas position equivalent error. The radar-absorbing materials damage dataset (RAMDD) is constructed by combining microwave images with optical images label. The semantic segmentation method is applied to filter out cluttering caused by scattering coupling or uneven distribution of absorbents. A new target detection model is proposed by changing training process, designing a composite backbone (CB) and adding model optimization methods (K-means clustering, cosine annealing, transfer learning and label smoothing) so that the model can detect the corresponding position of RAMs damage in the optical image via the microwave image rapidly. The mean average precision (mAP) of the target detection model is 76.04% on the RAMDD.