Research On Small Coil Counting Based On Fully Convolution Regression Neural Network

In recent years,with the widely discussion of the topic of artificial intelligence,the deep learning technology has gradually become a research hot in the Internet era.In the field of computer vision,it is more inclined to use deep learning technology to solve the problems of semantic segmentation,image classification and human posture recognition.Compared with traditional algorithms,its the advantage lies in the ability to learn features through a large number of samples to meet the target requirements.This paper mainly studies the visual counting problem of winding turns on micro terminals.In modern industry,the size of parts is getting smaller and smaller,and the quality requirements are getting higher and higher.In order to ensure the functional integrity of the whole part,quality inspections such as winding count and cross-coil identification must be performed after the winding is completed.Therefore,this article focuses on the above issues and the main contents are as follows:Firstly,a coil counting model based on image classification algorithm is proposed,and an intelligent counting system is developed.In the first place,the fast RCNN network is used to locate the coil around the column in the scene image,and a single coil around the column is separated and extracted as the sample set for subsequent counting research;Then the convolution neural network(CNN)model is used to train the coil samples with different turns to get the fixed class classifier.Finally,the coil turns are obtained by classification recognition,which avoids the low efficiency and high cost of manual counting.Secondly,a coil counting model based on Fully Convolutional Regression Neural Network(FCRN)is proposed.The model uses ASPP with different dilated rate to transform the Fully Convolutional Neural Network to construct multi-task network branches,namely semantic segmentation task branch and linear regression task branch,to obtain semantic segmentation prediction map and coil turn value respectively.This model achieves 89%of IoU in the private data set.It has a wide range of coil counts and is not limited by the type of coil samples,which effectively improves the adaptability and accuracy of counting.Thirdly,under the condition of zero sample of cross-coil,an algorithm(FCRN-DenseCRF)is proposed to modify the semantic segmentation prediction image of FCRN output by using the fully connected conditional random field to obtain more precise segmentation prediction image.After prediction image is projected horizontally,the cross coils can be identified according to the characteristics of the troughs of horizontal projection to improve the efficiency of standard coil counting.The two counting algorithms proposed in this paper are tested in the coil sample set,and can well complete the counting problem of winding turns.Compared with CNN classification counting method,algorithm of FCRN-DenseCRF can better solve the practical problems based on insufficient samples and crossed coil.It has strong robustness to complex winding turns and non-standard winding in factories,and has higher accuracy for standard coil counting.