Research On Marble Slab Cutting Based On Semantic Segmentation

The detection of irregular polygons with maximum enclosed rectangle is encountered in the process of leather cutting and the reuse of board scraps.At present,the method of adjudication is generally completed by manual estimation,This traditional way of manual inspection is extremely subjective,which not only causes waste of resources,but also is inefficient.Therefore,the factory urgently needs a kind of intelligent equipment to replace the artificial.Recently,the rapid development of depth learning and computer vision has provided some new ideas to solve these problems.To maximize the utilization of marble slab by algorithm,the main difficulties are two aspects: how to achieve real-time high-precision semantic segmentation of the marble slab image,and how to quickly and effectively find the largest enclosed rectangle of the target object.In this regard,the following experimental work has been done:Based on the semantic segmentation algorithm and utilize tensorflow & keras deep learning framework,this paper builds and trains an improved Fast-SCNN semantic segmentation model for real-time extraction of the marble slab image mask.Besides,this paper designs a maximum enclosed rectangle detection algorithm by the boundary point selection of the initial rectangle based on the center diffusion method,which realizes the rapid detection of maximum enclosed rectangle of marble plate.The detailed research works are as follows:1.Fully studied the current popular image segmentation techniques and compared the advantages and disadvantages between the methods based on traditional digital image processing and the methods by depth learning.2.In the aspect of semantic segmentation,after discussing the function of attention mechanism,By adding attention mechanism after upsampling high-level feature,the model has the ability to discriminat the importance of each channel.Studied the structure of Deeper atrous Spatial Pyramid Pooling module and improved the pyramid pooling module,make segmentation model can get more accurate segmentation image of edge details.Last,through a series of comparative experiments,the effectiveness and practicability of improved model are verified.3.In the aspect of maximum enclosed rectangle detection,several existing algorithms for detecting the maximum enclosed rectangle of irregular polygon binary images are researched,including traversing method,center diffusion method,traversal center diffusion method and genetic algorithm.Besides,designed a series of experiments to analyze the shortcomings of the existing methods of maximum enclosed rectangle detection and proposed an efficient and practical method: boundary-sorting growth method.4.With integrating all the above algorithms and experimental emulating whole algorithm at actual industrial scene,fast recommendation of maximum enclosed rectangle for marble slab is achieved and a suite of software has been developed accordingly.The experimental results show that the improved model achieves 97.35% pixel accuracy and 88.51% Mean Intersection over Union on marble slab datasets,respectively.the processing time of single image is about 70 Ms,it effectively realizes the real-time extraction of marble mask.The boundary ordering growth method proposed in this paper can fit an ideal maximum inline rectangle within 15 ms,effectively balancing the detection accuracy and speed.After the experiment and verification of the actual working condition,results show that overall algorithm in this paper achieves the maximum enclosed rectangle detection of marble slabs with an average time of 130 ms,it has great theoretical reference value and practical application value for the task of fast detection of the maximum enclosed rectangle of an irregular polygon object for leather,paper and board cutting.