An equal length log volume inspection system using deep-learning and Hough transformation

LIN Yaohai;ZHAO Honglu;YANG Zecan;LIN Mengting;College of Computer and Information Sciences,Fujian Agriculture and Forestry University;Key Laboratory of Smart Agriculture and Forestry (Fujian Agriculture and Forestry University);

log volume inspection;;Hough transform;;circle detection;;deep-learning;;object detection;;YOLOv3-tiny

Aiming at examining the end images of logs stacked in bundles,a volume inspection system is developed using machine vision methods. The robustness of the algorithm and the convenience of system application are two key problems. In order to realize the automatic inspection of log volume and solve the above problems,an equal length log volume inspection system using deep learning and Hough transformation was developed. Firstly,after the pre-processing of the training samples such as rotating the image,adjusting the photo exposure ratio,adding noise to the image and so on to improve the robustness of the model,the YOLOv3-tiny convolutional neural network was used to detect the image of the log end face,and then obtain the target area corresponding to each log end face. Secondly,in order to calculate the accurate diameter of the log end profile from the target area obtained by YOLOv3-tiny,the Hough transform circle detection algorithm was applied on the target area of the original image,after removing the noise edges in the center of the target area. Finally,an equal length log volume inspection system with a simple and convenient user interface was developed,which can output the volume of each log from the value of a real log diameter and the length entered by the user. It can realize the volume inspection of all logs in the image. The system was tested by many kinds of log end face images,including clean log end face image,log end face with some cut marks,log end face with some moldy,log end face image with complex environment,etc. The result showed that the algorithm proposed in this study had the true detection rate of 98.79%,and the false detection rate of 0.602%. The result confirmedthat the system had good robustness in many kinds of complex log end face images. The results also showed that the system was very robust on various complex log end face images. On the other hand,for compatibility with deep learning algorithm and user interface design,the operation interface of the system was implemented on the PyQt5 core library,which is simple and easy to use.