基于改进ResNet-18与高光谱特征的孵期种蛋性别无损鉴别方法

Non-destructive Sex Identification of Hatching Eggs during Incubation Based on Improved ResNet-18 and Hyperspectral Features

摘要:

针对传统种蛋性别鉴定方法存在的侵人性、效率低、依赖单一时间点数据等问题,为实现孵期种蛋性别无损、高精度检测,以京粉1号种鸡蛋为对象,提出了一种融合高光谱成像与深度学习方法,构建多阶段时序动态检测体系。自主研发双通道高光谱采集系统,于孵化第4、7、10、13天采集400~1000nm波段数据,通过椭圆拟合并提取蛋体中心感兴趣区域(ROI),经Savitzky-Golay平滑、主成分分析(PCA)降维及数据增强预处理后,构建引入通道注意力(SE)模块的改进ResNet-18模型,并结合长短时记忆(LSTM)模块实现多阶段时序特征融合。结果表明:孵化第10天为最佳检测窗口,单时期模型性别鉴定准确率达82.99%;时序融合模型准确率进一步提升至85.2%,较标准ResNet-18提升3.3个百分点;改进模型参数量仅9.9×10?,推理时间为47ms/样本,兼顾检测精度与效率。本研究提出的方法破解了传统技术的滞后性与侵人性缺陷,为家禽养殖智能化、绿色化发展提供了可靠的技术方案。

Abstract:

Aiming at the problems of invasiveness, low efficiency and reliance on single-time-point data in traditional hatching egg sex identification methods, to realize non-destructive and high-precision sex detection of hatching eggs during incubation, taking Jingfen No. 1 breeding eggs as the research object, a method integrating hyperspectral imaging and deep learning was proposed, and a multi-stage temporal dynamic detection system was constructed. A dual-channel hyperspectral acquisition system was independently developed to collect data in the 400~1000nm band on the 4th, 7th, 10th, and 13th days of incubation. The central region of interest (ROI) of the egg body was extracted via ellipse fitting. After preprocessing, including Savitzky-Golay smoothing, principal component analysis (PCA) dimensionality reduction and data augmentation, an improved ResNet-18 model incorporating the squeeze-and-excitation (SE) attention module was constructed, and multi-stage temporal feature fusion was realized by combining the long short-term memory (LSTM) module. The results showed that the 10th day of incubation was the optimal detection window, with the sex identification accuracy of the single-period model reaching 82.99%. The accuracy of the temporal fusion model was further improved to 85.2%, 3.3 percentage points higher than that of the standard ResNet-18. The improved model had only 9.9×10? parameters and an inference time of 47 ms per sample, balancing detection accuracy and efficiency. The proposed method overcame the lag and invasiveness defects of traditional technologies, and provided a reliable technical scheme for the intelligent and green development of poultry breeding.