Research On Detection Method Of Unsound Grains In Wheat Based On THz Imaging Technology

Wheat unsound grain are damaged wheat kernels that still have use value,including sprouted grains,moldy grains,insect-corroded grains and mutilated grains,etc.The content of wheat unsound grain is an important indicator of wheat quality,and the detection technology of wheat unsound grain is of great significance to improve wheat quality.Traditional unsound grain detection methods,such as manual sensory and machine vision,have problems such as cumbersome process,small amount of information,and inability to penetrate wheat unsound grain.Terahertz,as a new detection method,has the characteristics of high penetration and low energy,which provides a new method for rapid nondestructive detection of wheat unsound grain.However,the interference of noise and other factors in the process of acquiring THz images of wheat unsound grain leads to the problems of poor image quality and outstanding features in the original images.In this paper,we conduct in-depth research on THz image processing of wheat unsound grain,combine deep learning and width learning theories,propose THz image enhancement algorithms and detection algorithms for wheat unsound grain,improve THz image quality,highlight image details,achieve rapid and accurate detection of wheat unsound grain,and enrich and develop THz detection theories and methods.The main research contents are as follows:(1)The THz imaging system was used to acquire five types of wheat unsound grain samples,including normal wheat,sprouted wheat,moldy wheat,mutilated wheat and insecteaten wheat,to obtain the THz image data of wheat unsound grain,and to remove the background interference of the original wheat unsound grain THz image and highlight the information of the main contour of wheat unsound grain.The composition of THz images of different types of wheat unsound grain was analyzed to compare the differences of contours and features of different types of wheat.It provides a solid theoretical and data basis for the next experiments.(2)To address the problems of poor quality and outstanding edge detail features of the original wheat unsound grain THz image,the CBDNet-V wheat unsound grain THz image enhancement algorithm is proposed based on the CBDNet denoising algorithm and VGG feature extraction algorithm.The experimental results demonstrate that this model can effectively improve the image quality and highlight the image detail features compared with the traditional denoising algorithms of BM3 D,Dn CNN and FFDNet,and the PSNR and SSIM of the enhanced images are 39.267 d B and 0.942,respectively.CBDNet-V can effectively enhance the THz images of wheat unsound grain and improve the recognition accuracy.(3)For the problems that the traditional deep learning algorithm becomes more and more complex with the deepening of the model structure and the data processing time becomes longer,the width learning idea is introduced to build a THz image detection model of wheat unsound grain based on R-F-BLS.The RIDNet denoising module and FPN feature extraction module are embedded in the BLS network structure,and the CAM attention mechanism is introduced to improve the image quality and extract image features to achieve fast and accurate recognition of wheat unsound grain.The experimental results show that the model of R-F-BLS improves the network model accuracy and reduces the network complexity at the same time compared with deep learning algorithms such as CNN,Google Net and Res Net,and obtains 96.53% classification accuracy,which is 14.26%,9.26% and 7.84% compared with deep learning algorithms,respectively.