Research And Implementation Of Bamboo Chips Detection System Based On Improved YOLOv3 Network

China is the country with the largest output of bamboo products.Among them,bamboo mats have a wide market space as daily necessities.In the production of bamboo mats,there are phenomena such as poor appearance consistency.Traditional target detection methods are based on shallow features such as bamboo contours and textures for identification and screening,which has the problems of difficult parameter adjustment and low recognition accuracy.In response to these problems,based on the comparison of mainstream convolutional neural network detection algorithms,this paper selected YOLOv3,which has the fastest algorithm speed and high detection accuracy,as the basis of bamboo chips detection algorithm.By adapting the model training and detection recognition of the deep features of the bamboo,the adaptability problem of the traditional algorithm was solved,combing the lightweight improvement of neural network algorithm,the cost of computing hardware was reduced,and the calculation speed and accuracy were improved.This paper applied the improved algorithm to the bamboo chips detection system and achieved the expected results.The main work of the thesis was as follows:(1)In view of interferences in the bamboo samples collection process,this article compared and analyzed a variety of background segmentation algorithms and recognition algorithms,selected the random forest semantic segmentation algorithm with the best bamboo segmentation effect,as well as the YOLOv3 object detection algorithm that could effectively identify sticky bamboo segments,which laid a good foundation for the detection of bamboo chips.(2)Aiming at the time-consuming problem of manually labeling semantic segmentation data,a semi-automatic semantic segmentation method was proposed,which iteratively used operations such as sampling labeling,model training,and local calibration until the labeling results meet the requirements.This method effectively saved the strength of manual labeling.(3)For the problem of low adaptability of the detection algorithm model to different environments,this paper used the combined method of saturation,exposure,hue,rotation transformation and splicing transformation on the labeled bamboo dataset to enhance the data.The experimental results showed that the training model after data enhancement has strong environmental adaptability.(4)In order to improve the transfer efficiency of features in the network layer and reduce the cost of computing hardware,the YOLOv3 network was optimized and improved based on the characteristics of bamboo datasets.A dense series structure was used to replace the original residual connection structure,a convolution and maximum pooling combined dimension reduction structure was used to replace the original convolutional dimension reduction structure,and small-scale feature detection channels were eliminated..The experimental results showed that the total floating-point calculation amount,model size,and calculation time of the improved model were significantly reduced,and they could be loaded and run on GPUs with lower configurations.Experiments also showed that the improved model has good adaptability on Kaggle and Caltech open source datasets.(5)According to the requirements of bamboo chips production and application,the bamboo chips detection hardware and software control system were designed and implemented,and the improved algorithm was applied to bamboo chips detection.After on-site testing,the system had recall rates of more than 97% for all types of bamboo chips,and precision rates of more than 95%.The average detection speed is three times that of the traditional system,reaching 50 tablets per second.