Stacked-Sheets Image Segmentation With A Convolution Neural Network Algorithm

The precise counting of sheet-type products such as paper,panels and various types of cigarette labels is of great significance t o the cost control of related companies.However,due to the large number and variety of such sheet products,early manual counting and physical counting methods can't be used on a large scale due to low efficiency and large errors.Although traditional image processing algorithms increase the automation of stacked sheet counts,in the face of the common problems such as inconsistent thickness,adhesion,breakage and low contrast,this kind of algorithm can only use the single feature of the laminated image,so it can't achieve compatibility counting of multiple laminated slices.Aiming at these problems,a semantic segmentation method based on convolutional neural network that counts the connected regions of the laminated image to obtain the number of slices was designed in this paper.This method obtains more accurate segmentation results by obtaining larger context information,and then obtains more accurate number of laminated slices,which effectively solves many problems that have troubled traditional image processing algorithms.The specific content of this paper is as follows:(1)Aiming at the problems existing in the lamellar stack counting of traditional image processing and combining with the relevant theories and the characteristics of sheet laminate imaging,the scheme of first segmentation and then counting of laminar images by using the deep convolutional neural network method for the first time was proposed and implemented in this paper.(2)Combined with the characteristics of the stripe distribution of thin slices,the laminated end images were segmented by small image training and full image testing and combined with various data enhancement methods,the scale of the training set to be labeled was greatly reduced,making the training and testing of the model more efficient and feasible.(3)Based on the current situation that there were no relevant data sets available,a special data set is designed.The strip-shaped standard labels are created by using the slice center line,and the training model learns the sheet feature information in the center line label,thereby obtaining the semantic segmentation graph of the laminated image of the wafer for the final counting.(4)The classical u-net network was adjusted and improved to realize the accurate segmentation of thin slice images with fewer layers.The self-learning ability of the model was demonstrated through the middle activation of the model and the visualization of convolution kernel.Finally,the advantages of the proposed algorithm in laminated image counting are verified by comparing the laminated images with different thickness anomalies.That the comprehensive counting accuracy of the proposed algorithm is above 99% was proofed by many experiments,which is more robust than the traditional image algo rithm.