Cocoon Detection Algorithm Research Based On Deep Learning

With the year-on-year growth of domestic and foreign demand for silk products,the demand for raw material cocoon is also increasing.In the silk reeling production process,the cocoon selection process determines the quality and efficiency of reeling,and an efficient cocoon selection method can reduce the loss of raw cocoons and energy consumption.Traditional cocoon selection process is divided into coarse and select,coarse selection process is traditionally used to select cocoons by the human eye,that is,according to the appearance of cocoon characteristics or cocoon weight,by the human eye to identify a variety of different types of raw materials cocoon screening method.This manual screening method is influenced by the operator’s subjective factors,which can easily lead to low recognition rate,unstable accuracy and other problems,can not meet the requirements of modern factories batch,efficient and accurate production,but also does not meet the development trend of intelligent manufacturing in the textile industry.Meanwhile,deep learning and machine vision technologies have made rapid progress in the recent past.Therefore,the use of contemporary advanced deep learning technology for the cocoon selection process of silkworm cocoon intelligent,automated transformation,both feasible and will be important to promote the development of China’s cocoon silk industry.The main research contents of this topic are as follows:1.build a large scale,suitable for training deep neural network cocoon image dataset.This part mainly includes:the development of data set acquisition and production plan;the acquisition of large batch cocoon image data;cocoon image annotation.2.On the cocoon image dataset constructed by ourselves,the current typical deep learning target detection frameworks(Faster R-CNN,Yolo v3,SSD)are trained,their performance parameters indexes are compared and analyzed,and the optimal automated cocoon selection model is selected from them.And based on this,the impact of different image data augmentation methods on the cocoon detection model is explored from the perspective of the dataset,so as to further improve the generalization ability and robustness of the model.3.a new cocoon selection model C-SSD is proposed based on the selected deep learning cocoon selection model.this part mainly includes:the best way to connect the attention mechanism is derived through classification comparison experiments,and this attention mechanism is introduced into the selected cocoon selection model,so as to improve the cocoon selection accuracy of the new model.The results of the above study found that:1.three typical target detection frameworks(Faster R-CNN,Yolo v3,and SSD)trained on self-developed cocoon datasets have detection accuracies of 0.964,0.940,and 0.924,and detection rates(fps)of 1fps,58fps,and 72fps,respectively.in terms of accuracy,Faster R-CNN is the highest,and SSD is the fastest in terms of detection rate.The recall rates of the three models for good cocoon were 0.92,0.90,and 0.81.Since the target detection algorithm should not only meet the requirements of detection accuracy,but also meet the requirements of real-time processing of data frames,therefore,in a comprehensive view,the SSD model has a high application value in the automated cocoon selection process.2.by training the SSD model with different data augmentation modes,it is found that when the content information of the cocoon image is not changed and only the morphological transformation of the image is performed,it has a better effect on improving the generalization ability and robustness of the model.3.In the experiment of improving the SSD model,the detection accuracy of the C-SSD algorithm for cocoons after introducing the attention mechanism is improved from 0.940 to 0.950.