YOLOSpecNN: A novel γ-ray spectra full-energy peak automatic search and segmentation model inspired by YOLO

摘要: The qualitative identification and quantitative analysis of radioactive nuclides in unknown environments are essential for remote monitoring and prompt early warning of radioactive contamination. In recent years, deep learning techniques have made significant strides in automated qualitative identification. However, the quantitative analysis of radioactive nuclides still depends on traditional methods to determine peak positions and boundaries. These methods often require extensive manual expertise and parameter tuning, which makes it challenging to meet the demands of unmanned remote monitoring. This paper presents a novel framework for automatic full-energy peak segmentation, named YOLOSpecNN. We introduced a multi-Root Mean Square Error (RMSE) joint optimization function and developed a unified regression model capable of simultaneously predicting the central position, boundaries, and confidence of full-energy peaks. To address the challenge of low recall rates due to narrow, weak, and overlapping peaks, we proposed a new multi-scale context feature extraction module (MSNN module). This module effectively enhanced local detail features, significantly improving recall rates. The effectiveness of the proposed method was validated using six artificial radioactive nuclides (241Am,57Co,131I,134Cs,137Cs,and 60Co), along with 40K, to construct a mixed energy spectrum dataset for quantitative evaluation. Experimental results show that the proposed method significantly outperforms traditional approaches, achieving a precision of 0.998, recall of 0.95, and the best F1 score of 0.974@0.427, and the average precision of 0.946. Compared to traditional morphological methods, the proposed method improves precision, recall, and the best F1 score by 0.512, 0.199, and 0.391, respectively. Ablation experiments further reveal that the MSNN module notably enhances recall, with an improvement of 0.067. Moreover, the proposed method performs excellently even in challenging environments with low gross counts and low Signal-to-Noise Ratio (SNR), achieving state-of-the-art (SOTA) results. Additionally, the model achieves an average real-time inference performance of 16.1941 ms on a 15 W low-power device. Overall, the proposed method demonstrates exceptional performance in the automatic search and segmentation of full-energy peaks, offering robust support for the implementation of unmanned remote radiation monitoring systems.