Research On Wheat Impurity Content Detection Based On Cone Beam CT Images

Grain quality and safety are the main concerns related to national security.In recent years,frequent natural disasters in China have had serious impacts on both the quantity and quality of grain production.The COVID-19 pandemic has also led to restrictions on the export of grain in many countries worldwide,making it particularly important to ensure an adequate and high-quality grain reserve.In addition,microorganisms and insect eggs in impurities can easily cause grain storage pests,mold,or heating,which can affect grain storage safety.Currently,most methods for measuring impurity content in grain are still at the stage of manual screening and calculation,with high levels of subjectivity in the results.Therefore,an efficient and accurate method for detecting and calculating impurity content in grain is of great significance for grain quality evaluation and safety.In this study,the research focus on detecting impurities in piled wheat,using cone-beam computed tomography(CT)scanning to collect wheat and impurity images.This study proposes CT image filtering algorithms to improve the quality of wheat sample images,establish a CT image dataset of wheat samples with impurities,and then construct a rotation box object detection network model based on the YOLOv5 object detection network to detect wheat impurities.Based on this,this study use the two-dimensional CT slice images of wheat samples to reconstruct them into three-dimensional voxel images,enabling the calculation of wheat impurity content.The research in this paper mainly includes:(1)In response to the interference such as noise and artifacts present in CT images obtained from scanned wheat samples,this research investigates suitable filtering methods for wheat sample images and establishes an image dataset of wheat samples with impurities.Due to the high requirements for image reconstruction quality in wheat impurity content detection,it is necessary to filter the data obtained from cone-beam CT scanning to reduce the influence of oscillation effects and preserve image details.This study employs a hybrid filtering method to process CT images,which first uses the damping properties of the S-L filter to suppress high-frequency components in the projection image and alleviate oscillation phenomena.At the same time,the Blackman filter is used to sharpen the image and preserve more image detail components.Finally,the CT images are encoded to establish an image dataset of wheat samples with impurities.(2)To address the problem of geometric mismatch between the rectangular anchor detection windows used by traditional object detection algorithms and the shape of impurities in wheat sample images,this study developed a rotation box-based object detection network model based on the YOLOv5 object detection network.The traditional horizontal annotated boxes used in YOLOv5 were replaced with rotationally annotated boxes,which reduced the amount of annotation work required while also reducing the invalid image area within the selected region,thereby reducing the complexity of the object detection network.The traditional label representation was also changed to a Circular Smooth Label(CSL)with angle information,which better matches the geometric shape of impurities and solves the problem of boundary loss caused by angle regression.In the loss function,the CIo U function was used instead of the original GIo U function to reduce the loss value and achieve stable regression of the boundary.Experimental results show that the model achieved an average precision of 88.83% for impurity recognition in the wheat sample dataset,with an average detection time of 11.6ms.(3)In view of the significant subjectivity in traditional wheat impurity detection methods,a three-dimensional voxel-based approach for calculating the wheat impurity content is proposed in this study.Firstly,the filtered images are segmented using a region growing algorithm based on grayscale values to separate different materials.Then,the segmented two-dimensional images are reconstructed into a three-dimensional voxel-based model using a combination of volume rendering and surface rendering.Next,the impurity volume is calculated based on the reconstructed three-dimensional model,followed by the introduction of the impurity density and volume formulas to achieve the calculation of the wheat impurity content.The experimental results show that the relative error between the impurity content calculated by the proposed method and the actual impurity content is0.22%.