基于可见光图像的水稻颖花开花状态检测方法

Method for detecting rice flowering spikelets using visible light images

通过识别水稻开花张开颖花内外颖与吐出颖花花药的特征,进而准确判断颖花开花时间,是及时进行杂交水稻制种授粉的前提。该研究通过可见光相机获取水稻颖花图像,基于可见光蓝色通道串联大津法(Series Otsu,SOtsu)提取颖花花药,同时使用深度学习算法基于区域的快速卷积神经网络(Faster Regional Convolutional Neural Network,FasterRCNN)及YOLO-v3识别颖花花药与张开颖花内外颖,通过对比不同算法识别精确率、召回率、F1系数以及皮尔逊相关性系数,研究适用于识别颖花开花状态的特征与方法。结果显示,FasterRCNN算法检测张开颖花内外颖精确率达1,召回率达0.97,F1系数为0.98,皮尔逊相关系数为0.993,串联大津法检测吐出花药精确率达0.92,召回率达0.93,F1系数为0.93,皮尔逊相关系数为0.936。这表明串联大津法与FasterRCNN算法适用于水稻颖花开花状态检测,且张开颖花内外颖比吐出花药更适合作为水稻开花状态特征应用于深度学习算法检测。串联大津法可代替FasterRCNN算法在模型构建完成前进行检测,保证水稻颖花开花状态检测连续性。

Rice flowering spikelets bloom generally at 10:00-12:00,especially when the temperature is 24-35℃and the relative humidity is 70%-90%.Therefore,the flowering time is necessary to be accurately determined for the timely pollination in the production of hybrid rice seed.In this study,the images were captured by a visible light camera at two flowering characteristics,including the opening of spikelet hull,and the emesis of spikelet anthers.Series Otsu(SOtsu)was applied in tandem to extract the spikelet anthers through the visible light blue channel.An attempt was made to detect the flowering status of rice glumes using visible images,in order to meet the needs of hybrid rice seed pollination.A Canon single-lens reflex(SLR)camera was adopted for data acquisition,which was a benefit to segment the image using the tandem SOtsu.Deep learning models,such as FasterRCNN and YOLO-v3,were used to identify the spikelet anthers and the opening spikelet hull.The most suitable method was selected for flowering characteristics detection to compare the precision,recall,and the F1 coefficient of different models.Two datasets of visible light images were set for spikelets(15 cm and 45 cm imaging distance),each of which used two characteristics.A labeling software was applied to label the category and position of images.As such,a sample database was established for the training of detection models with deep learning.The performance of three models,including SOtsu,FasterRCNN,and YOLO-v3,were evaluated,where the detection was verified from multiple angles.The experiment was also conducted for the model robustness as well.The maximum inter-class variance was utilized in the SOtsu to separate the foreground(rice)from the background using the grayscale image of B-channel,where the grayscale of the background was set to be zero.An analysis was then made for the maximum inter-class variance that applied independently in the pixel range of extracted region,and then the spikelet anthers were further separated from the spikelets hull.The original gray values of spikelet anthers were retained,while the gray values of spikelet hull were set to be zero.Finally,the extraction was evaluated to combine with original images and the number of connected areas that were calculated by the eight-connected output as well.The results showed that the precision,recall rate,F1 coefficient and Pearson correlation coefficient of FasterRCNN model in spikelet hull detection were 1,0.97,0.98,and 0.993,respectively,while those of SOtsu in spikelet anthers detection were 0.92,0.93,0.93,and 0.936,respectively.It inferred that the SOtsu and FasterRCNN models were both capable of rice flowering detection,but the opening spikelet hull was more suitable than the spikelet anthers for the rice flowering features detection with deep learning model.The model robustness indicated that the highest stability was achieved in the FasterRCNN model to identify the spikelet flowering status with high precision under low,high and uneven light conditions.In addition,the spikelet anthers that opened on the same day split and pollened in 3-5 min,and withered on the same day.There was no recognition significance in the withered spikelet anthers without pollen.It was also necessary to verify the classification ability of detection for the withered anthers,in order to avoid the wrong identification of withered anthers.The SOtsu performed well in the image segmentation using the gray value for the withered spikelets anthers.The SOtsu was also better than FasterRCNN in the identification of withered spikelets.Correspondingly,the SOtsu was expected to replace the FasterRCNN model for the flowering spikelet before the completion of model construction,in order to ensure the detection continuity of rice flowering spikelet.The influencing factors of recognition were reduced to control the process of detection.Since the segmentation was processed for morphological opening operations,there was also some limitation in the recognition of overlapping anthers.A further study can be followed by a more in-depth exploration of high-throughput detection.