| In the production of industrial and traditional Chinese medicine preparations,particle detection is a frequently encountered detection requirement in the classification and processing of raw materials.In this paper,we use sand commonly used in construction as a typical representative of irregular particles,and adopt a variety of classical image processing and deep learning image recognition methods to achieve segmentation,classification and statistical comparison of irregular particle images.Classical image processing algorithms are widely used in industrial inspection.Because the industrial detection environment is relatively stable and the detection content is relatively simple,the classical algorithm can realize the detection efficiently.However,the classical image processing algorithms have high requirements on parameter settings when performing detection tasks,and have certain requirements on the position,shape,and color of the target.In this paper,when detecting irregular particles of sand,the threshold,edge segmentation and watershed algorithms are prone to false detections and missed detections.For this reason,we further carry out research on detection algorithms based on deep learning.Deep learning algorithms have carried out a lot of research in the field of image detection,such as autonomous driving,product defect detection,medical image segmentation,etc.The research on deep learning algorithms is mainly divided into two aspects,one is to study how deep learning networks can extract image features more effectively,and the other is to study more effective model frameworks for different detection tasks.In this paper,the Mask R-CNN instance segmentation model is selected,and feature extraction networks with different structures are designed to detect sand particles,and the detection information obtained by the model inference is optimized and adjusted by the classical algorithm.In this paper,an image processing method is proposed to optimize the mask predicted by the model.The method first uses the prediction box output by Mask R-CNN to select the region of interest,and then uses the image segmentation method to determine the candidate mask in the region of interest.The candidate mask contains the sand itself and non-sand masks such as the background,and then removes the non-sand mask parts according to the predicted mask output by the model,and finally obtains the accurate output of the sand mask.Compared with the original model,the method proposed in this paper can improve the accuracy of the mask and effectively verify the particle size detection.In addition,in order to use the low-level feature layer information more efficiently,this paper selects and compares different feature extraction networks,and compares the use of Res Net,Dense Net,Efficient Net and Dense Net enhanced by the channel attention mechanism.The channel attention mechanism is an important method to improve the feature extraction capability of the convolutional network.The attention mechanism module is added to the Dense Net network module.Compared with the detection accuracy that Dense Net can increase,the smaller feature layer of Dense Net is used to achieve equivalent feature extraction ability.In this paper,the detection method of sand particles is discussed,and the deep learning and image processing methods are organically integrated.The results show that the method in this paper can have high accuracy in sand particle detection and size calculation.In addition,the method in this paper is also applicable to the detection of other irregular particles,and has certain reference significance and value for the detection and quality monitoring of other irregular objects such as cells,other mineral particles,and artificial sand. |
