基于热图像的种蛋气室变化监测算法

种蛋气室的大小是监测种蛋孵化过程的重要指标之一。根据种蛋的热力学结构，种蛋在孵化过程中，包裹气室部分蛋壳会与其他部分蛋壳产生温差，从而可通过热红外图像进行观察。针对在种蛋孵化过程中，人工照蛋检测气室效率低的问题，探索设计了一种基于热图像的种蛋气室变化俯视监测算法。监测种蛋气室热图像的算法主要包括种蛋目标检测，种蛋图像分割和种蛋气室面积计算3个部分，其中种蛋的目标检测采用Faster-RCNN算法实现；种蛋图像分割采用BP神经网络算法实现；种蛋气室面积是在种蛋图像分割的基础上进行计算。使用孵化5天及以上的种蛋作为研究对象，并拍取种蛋的热图像进行试验。试验结果表明：种蛋热图像的目标检测的平均精度（mAP）为99.85%，拥有较好的检测效果。使用BP网络对种蛋进行图像分割。BP神经网络经过调参后，其网络最佳的结构为三层隐藏层，每个隐藏层拥有1 000个神经元，最优初始学习率为0.000 1，最优最大迭代次数为500。以F1-measure作为分割效果的评价指标，BP神经网络的图像分割总体结果为87.02%，Otsu算法的总体结果为65.25%。其中只有一个蛋的情况下，BP神经网络的分割结果为87.17%，Otsu算法的结果为68.86%。存在其他种蛋的干扰条件下，BP神经网络的分割结果为86.94%，Otsu算法的结果为61.64%，BP神经网络的分割效果优于Otsu分割算法，BP神经网络拥有更强的抗干扰能力。最后提取了孵化5～19 d种蛋的气室变化，通过观察种蛋气室大小曲线来监测种蛋的孵化情况，可看出随着天数的增加，气室有着明显变大的趋势。人工测量法与热红外测量法比较结果说明两者相关性为0.934 3，拥有较好的相关性。基于热图像的种蛋气室变化监测算法可在实际生产中实现种蛋的识别与气室大小的快速监测，为实现监测种蛋孵化的自动化提供了技术参考。

Fertilized Eggs’Air-Cell Change Monitoring Algorithm Based on Thermal-Image

The size of the egg air-cell is one of the important indicators for monitoring the hatching process of the eggs.According to the thermodynamic structure of the breeding egg,during the hatching process of them,the temperature difference between the part of the shell enclosing the air-cell and the other part of the shell may cause a temperature difference,which can be observed by thermal infrared imaging technology.A thermal-image based monitoring method for egg air-cell change was designed.The algorithm for monitoring the thermal-image of the egg air-cell mainly includes three parts:egg target detection and segment,egg air-cell’s size of segment and egg air-cell’s area calculation.The target detection of the eggs is implemented by the fast-RCNN algorithm.The size of the egg air-cell is implemented by BP neural network.The egg air-cell area is calculated based on the segmented egg thermal-image segment.In this paper,eggs had hatched for 5 days or more were used as research objects,and thermal-images of them were taken for testing.The test results show that the mean average precision(mAP)of the target detection of the thermal-image for the egg is 99.85%,which has a good detection effect.Under optimizing of the BP neural network’s hyperparameters,the result shows that the best structure of layers is 1000-1000-1000,the best initial learning rate is 0.0001,and the best max-iteration is 500.Using F1-measure as the evaluation index of the segment effect compare with the Otsu algorithm,the BP neural network’s result is much better than the Otsu algorithm.The Otsu algorithm’s segment evaluation is 65.25%,and the BP neural network’s result is 87.02%.In the case of only one egg,the segment result of the BP neural network is 87.17%,and the result of the Otsu algorithm is 68.86%.The segment result of BP neural network is 86.94%,and the result of the Otsu algorithm is 61.64%under the interference of other eggs.BP neural network has a stronger anti-interference ability.At the end of the experiment,the air-cell changes of fertilized eggs from 5 to 19 days were extracted,and the hatching of the eggs was monitored by observing the curve of the egg chamber area.The curve shows that the air-cell tends to become larger with the days increasing.The comparison between the artificial measurement method and the thermal-image measurement method shows that the correlation between the two is 0.9343,which has a good correlation.The thermal-image of egg air-cell change monitoring algorithm can realize individual identification of egg and rapid monitoring of gas chamber size in actual production,which is of great significance for health monitoring during egg hatching.