Research On Detection Of Unsound Wheat Kernels Based On Instance Segmentation

The assessment of wheat quality is crucial during its market circulation.The quantity of defective kernels per unit volume of wheat is a key factor influencing the quality evaluation.Therefore,identifying defective kernels holds immense significance for ensuring high-quality wheat.However,in practical scenarios,assessing a batch of wheat samples for defective grain content requires manual screening by warehouse personnel,which incurs substantial labor and time costs.The present machine vision technology exhibits a low degree of accuracy in detecting and recognizing adhesive grains,thereby making it arduous to identify them.To address the aforementioned issues,this paper focuses on the identification of perfect grains and unsound kernels in wheat,including but not limited to moldy kernels,injured kernels,spotted kernels,sprouted kernels,and broken kernels.In order to achieve this objective,a wheat data set is created.After comparing the experimental outcomes of various models,we opted for and established an unsound wheat kernel instance segmentation model using an enhanced Mask RCNN.With the segmentation outcomes,we devised a regression model to measure the wheat grade by computing its area and mass.Ultimately,we accomplished the development of a visual platform for detecting unsound grains.The precise tasks performed in this paper are as follows:(1)Create a dataset for wheat.Due to the current absence of a comprehensive wheat grain dataset,we acquired 130 images,comprising a total of 26,754 grains,through our established image acquisition platform to serve as the primary dataset.The grain was meticulously labeled using labelme software,resulting in corresponding annotation files and mask images.To enhance the model’s generalization capability,the image’s local features and intriguing attributes are deliberately emphasized.Additionally,the original dataset is fortified through five data augmentation techniques: brightness reduction,noise amplification,random point addition,translation,and flipping.After being divided in an 8:1:1 ratio,the data set was used to construct 780 COCO pieces of wheat grain data.(2)The study involved an investigation into the fundamental model for detecting and segmenting unsound wheat grains.To ensure the practicality of wheat unsound grain segmentation,various methods such as target detection,semantic segmentation,and instance segmentation were analyzed and compared.Through experimentation,the instance segmentation method was found to have superior detection accuracy and was more suitable for post-processing rating.After comparing several classical case segmentation models,the Mask RCNN model was chosen due to its higher segmentation accuracy and better ability to segment adhesion particles.Preliminary experiments confirmed that the basic Mask RCNN model could effectively recognize and segment unsound wheat grains.(3)An improved Mask RCNN unsound grain detection model was proposed.In order to solve the problem that the backbone is not targeted and inaccurate when extracting features,efficient channel attention(ECA)module is added to the four feature layers of the backbone feature extraction network,respectively,to improve the adaptive ability of the model for different types of wheat particles.In order to enhance the network’s extraction of wheat edge information and adapt to the size changes of different wheat grains,a bottom-up feature pyramid network was supplemented after the backbone feature network so that the abstractlevel features extracted from the network could integrate more bottom-up location information and edge information,which was conducive to localization and segmentation.In the process of extracting the area of interest and generating candidate boxes in the original Mask RCNN network,many anchor boxes need to be generated and a lot of position information needs to be saved,which increases the amount of computation of the model and increases the inference speed and memory consumption of the model.In view of this,the region proposal network(RPN)is adjusted.At the same Time,the non max suppression(NMS)strategy is optimized in the post-processing stage to prevent the model from identifying the highly cohesive particles as the same target.The experiment shows that the m AP of the improved model is increased from 47.2% to 86.3% and the processing Time of an image(about 200 grains)is reduced from11 s to 8s.(4)A regression model for grain area and bulk density of wheat was developed,along with a visualization platform.The main objective of detecting imperfect granules is to determine their content,which is essential for grade evaluation.The calculation of imperfect grain content relies on bulk density;hence,a regression model was established to convert the segmented image’s area into bulk density.To further validate the regression model,ten batches of wheat containing various levels of imperfect grain were selected,and the model was corrected based on actual weighing results.Finally,Python’s Flask technology was employed to integrate the front and back ends and create an unsound kernel detection model.