| During the cocoons processing,it is necessary to eliminate the inferior cocoons that are unattainable or difficult to be reeling according to the requirements of silk making technology for improvement of the quality of cocoons silk.However,the current inferior cocoons detection mainly depends on manual visual inspection,which limits the objective evaluation and efficient detection of inferior cocoons.Aiming at the above problems,a method adopt machine vision detection instead of manual detection inferior cocoons is proposed in this thesis.The main research contents are as follows:(1)The image acquisition system of inferior cocoons detection is built and the inferior cocoons detection data set is established.Firstly,we choose the software and hardware platform of inferior cocoons image acquisition system reasonably;Secondly,appropriate shooting distance is selected for the linear scanning camera according to the field depth of the image acquisition system,and the parameters of the image acquisition system are further configured by calculating the sampling frequency.Finally,the collected linear array images are synthesized into area array images,and and the images are expanded and annotated to construct the cocoon detection dataset.(2)A real-time inferior cocoons detection model based on model channel pruning and receptive field enhancement is designed.Based on the YOLOv3 target detection model,the YOLOv3 model parameters are preset by getting the anchor suitable for inferior cocoons detection through K-means cluster analysis to improve the model accuracy;according to a preset pruning rate,the channel after sparse trained model is pruned based on batch normalization layer scaling factor to compress the size of the model;the receptive field block is embedded in the pruned model to enlarge the receptive field and enhance the discriminability and robustness of the model.The model weight of final YOLOv3-MC-RFB inferior cocoons real-time detection model is 52.10 M,the mean average detection speed is 48.86 frames/s,and the mean average detection precision is 94.08%.Compared with the original YOLOv3 target detection model,the parameters are com-pressed by 77.74%,the mean average detection speed is increased by 19.18 frames/s,and the mean average detection precision is increased by 6.60%.(3)A real-time inferior cocoons detection model based on light manipulation network is designed.Based on YOLOv4 target detection model,the candidate anchor parameters are preset by performming cluster analysis on the data set of inferior cocoons through K-means algorithm to improve the model accuracy.By adopting the method of model depth manipulation,the model is compressed to achieve lightweight and fast detection speed.In addition,the lightweight convolution module is designed for a lightweight feature extraction network to further improve the detection speed.The model weight of final YOLOv4-DM-LCM inferior cocoons real-time detection model is 145.00 M,the mean average detection speed is 49.37 frames/s,and the mean average detection precision is 95.55%.Compared with the original YOLOv4 target detection model,the parameters are com-pressed by 40.82%,the mean average detection speed is increased by 23.61 frames/s,and the mean average detection precision is increased by 1.87%.(4)Embedded platform deployment of inferior cocoons real-time detection models.The YOLOv3-MC-RFB and YOLOv4-DM-LCM inferior cocoons real-time detection models are deployed and tested on the Jetson Nano B01 embedded experimental platform.By comparing and analyzing the detection effects of YOLOv3-MC-RFB and YOLOv4-DM-LCM on the embedded experimental platform,YOLOv4-DM-LCM model is selected as the deep learning model to complete the inferior cocoons real-time detection task.This thesis provides the design basis of image acquisition system for inferior cocoons real-time detection equipment,and effective deep learning algorithms for real-time detection of yellow spotted cocoon,crushed cocoon,thin shelled cocoon and little cocoon is proposed. |
PDF Full Text Request