Developing thermal infrared de-ghost and multi-level nested conglutinated segmentation algorithm for detection of rice seed setting rate
Graphical abstract
Introduction
Rice is one of the most important food crops in the world, especially in Asia (Bellis et al., 2022, Wang et al., 2021). The quality and number of grains per panicle has a critical impact on rice yield and precision agricultural management (Lu et al., 2022). Rice grains can be divided into full grains and empty grains according to their fullness. In agronomy, grains with grouting degree greater than one-third are full grains, and grains with grouting degree less than one-third are empty grains. Full grains are the real component of rice yield. The seed setting rate (SSR) is defined as the percentage of full grains in all rice grains per panicle. It is an important phenotypic parameter to characterize rice yield (Zhang, 2007). Li et al. (2013) had reported that a RING-type E3 ubiquitin ligase could regulate the SSR. Xiang et al. (2019) found a new rice gene that could regulate the SSR by facilitating rice fertilization. In these breeding studies, it is necessary to calculate the SSR of experimental samples. So rapid and accurate detection of SSR is of great significance for breeding, phenotypic measurement and yield prediction (Gong et al., 2018, Guo et al., 2021).
Traditional measurement methods mainly relied on manual identification of grain plumpness, but manual visual method is time-consuming and labor-intensive(Guo et al., 2021). With the rapid development of breeding technology, hundreds of new breeding materials need to be processed every day. Therefore, there is an urgent need to develop a fast and reliable SSR measurement technology to break through the bottleneck of traditional grain phenotype measurement methods (Jiang et al., 2022, Li et al., 2022).
At present, plant phenotypic measurement technology which helps to increase the efficiency of breeding and improved cultivation has been a hot research area. Many researches mainly focused on the quality detection and grain classification by morphological, color and texture characteristics. Paliwal et al. (2004) used the color camera to collect grain images and applied them to classify wheat varieties and to identify moldy or abnormal grains. Lin et al. (2018) proposed a machine vision system based on deep convolution neural network (DCNN) structure. Compared with traditional methods, this system could improve the classification accuracy of three different groups of rice images.
However, there is little difference between the external features of full and empty grains. Therefore, it is difficult to complete the recognition of two kinds of grains and detect the SSR by image processing methods based on color and texture features. So many scholars have developed methods based on X-ray imaging and computed tomography (CT) to study the internal quality of grains. Karunakaran et al. (2004) applied soft X-ray imaging technology to grain internal quality detection. Their team also used near-infrared camera and carbon dioxide sensor to detect internal quality of grain. But these devices and techniques could not distinguish between full and empty grains. The solution of Duan et al. (2011) was that the fusion between two images obtained based on X-ray and line array camera could more accurately identify full and empty grains, but the low efficiency and high cost of X-ray made it difficult to popularize. Kong and Chen (2021)proposed a feature extraction and three-dimensional recognition method based on mask region convolution neural network (Mask R-CNN) (He et al., 2017) for rice panicle CT images. Combining CT feature characterization with Mask R-CNN algorithm, feature extraction and classification were carried out for panicles and grains of each layer of CT sequence to calculate the SSR. The difficulty and high cost of CT image acquisition, this method was not practical (Zhang and Xia, 2021).
In recent years, thermal infrared technology has developed rapidly, and the thermal infrared sensor with gradually reduced cost has become more and more popular in the detection of agricultural products. According to the temperature difference, thermal infrared technology can detect the internal quality of products. External information of agricultural products can be detected based on RGB image. Fusing thermal infrared image (TII) and RGB image (RGBI) to detect seed setting rate is a solution. Our research was aimed at exploring an automatic and accurate method to detect SSR by developing thermal infrared de-ghost, image registration and multi-layer nested conglutinated segmentation algorithm. The main objectives were: (1) to build a thermal infrared–visible light dual imaging system, and to collect TIIs and RGBIs of rice grains; (2) to propose three kinds of methods (thermal infrared de-ghost, image registration and deep learning) to detect the SSR, and to compare the performance of different methods; (3) to evaluate detection performance of the models.
Section snippets
Dual imaging system
A dual imaging system, called thermal infrared–visible light dual imaging system, was built for rice grain data acquisition (Fig. 1). A thermal infrared camera (FLIR tau2, 614 × 512 pixels, FLIR Systems Inc., Oregon, USA) and a RGB camera (Sony α 6000, 6000 × 4000 pixels, Sony Inc., Tokyo, Japan) were selected to build the sensor system with image acquisition and data storage functions. The system signal control part was composed of remote controller, receiver, signal trigger switch and
Grain temperature variation law
The FLIR tau2, which had completed the temperature measurement calibration when leaving the factory, could output 14-bit lossless raw format digital image. TII was scanned pixel by pixel to obtain the minimum gray value (Fmin) and the maximum gray value (Fmax). In order to facilitate the post-processing of the image, it was extended to 16-bit digital image g(x, y). The temperature calibration formula was used to obtain the temperature distribution diagram t(x, y) from original TII. The formula
Prediction performance analysis
To obtain the best SSR detection performance, two new detection methods of SSR were proposed in this study. The ASSR of image registration method combining RGBI and TII was 80.83%, and the ASSR of the thermal infrared de-ghost method was 97.66%. Compared with the three deep learning models (96.75%), the thermal infrared de-ghost method combined with multi-layer nested conglutinated segmentation algorithm proposed in this study achieved the best performance. Due to the difficulty of detection,
Conclusion
A dual imaging system was set up to collect two kinds of image information at the same time in this study. Compared with the three deep learning models and image registration methods, the highest SSR detection accuracy (97.6%) was achieved by combining thermal infrared de-ghost and multi-layer nested conglutinated segmentation algorithm. Our findings suggest that the method of fusing thermal infrared de-ghost with multi-layer nested condensed segmentation algorithm by combining RGB image (RGBI)
Funding
This work was supported by National Key Research and Development Program of China (2019YFE0125500-02), Science and Technology Department of Zhejiang Province (2022C02034, 2020C02016), and Collaborative Extension Program of Major Agricultural Technologies of Zhejiang Province of China (2021XTTGLY0104).
CRediT authorship contribution statement
Jun Zhou: Conceptualization, Methodology, Software, Formal analysis, Validation, Investigation, Writing – original draft, Writing – review & editing, Visualization. Xiangyu Lu: Methodology, Software, Validation, Investigation, Visualization. Rui Yang: Validation, Visualization. Yaliang Wang: Resources, Data curation. Huizhe Chen: Resources, Data curation. Jianxun Shen: Resources, Data curation. Mengyuan Chen: Software, Visualization. Zhenjiang Zhou: Writing – review & editing, Project
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (30)
- et al.
A deep learning framework for unsupervised affine and deformable image registration
Med. Image Anal.
(2019) - et al.
Fast discrimination and counting of filled/unfilled rice spikelets based on bi-modal imaging
Comput. Electron. Agric.
(2011) - et al.
Detection of internal wheat seed infestation by rhyzopertha dominica using x-ray imaging
J. Stored Prod. Res.
(2004) - et al.
Image registration using BP-sift
J. Vis. Commun. Image Represent.
(2013) - et al.
Detecting intra-field variation in rice yield with unmanned aerial vehicle imagery and deep learning
Front. Plant Sci.
(2022) - et al.
Image-based on-panicle rice oryza sativa l. Grain counting with a prior edge wavelet correction model
Agronomy.
(2018) - et al.
Automatic and accurate calculation of rice seed setting rate based on image segmentation and deep learning
Front. Plant Sci.
(2021) - He, K.M., Gkioxari, G., Dollar, P., Girshick, R. 2017. Mask R-CNN. 16th IEEE International Conference on Computer...
- et al.
Saliency-based YOLO for single target detection
Knowl. Inf. Syst.
(2021) - et al.
Drying temperature regulates vigor of high moisture rice seeds via involvement in phytohormone, ros, and relevant gene expression
J. Sci. Food Agric.
(2021)
Evidence and impacts of nanoplastic accumulation on crop grains
Adv. Sci.
Effects of quality characteristics on milled rice produced under different milling conditions
J. Korean Soc. Appl. Biol. Chem.
Mask R-CNN-based feature extraction and three-dimensional recognition of rice panicle CT images
Plant Direct.
Brown rice germ integrity identification based on deep learning network
J. Food Qual.
Natural variation in ptb1 regulates rice seed setting rate by controlling pollen tube growth
Nat. Commun.
Cited by (1)
Method of Peanut Pod Quality Detection Based on Improved ResNet
2023, Agriculture (Switzerland)