A coil counting model based on full convolution regression neural networks

Abstract:
This paper studies the visual counting problem of the number of winding turns on the micro terminal. Its main task is to distinguish the counting of cross coils and standard coils. Relying on manual counting is not only inefficient but also has large error. In order to improve the work efficiency of product detection, deep learning technology is used to realize automatic counting. The common deep learning counting model is realized by convolution neural network classification, but it can only recognize the fixed number of turns. A full revolution region neural network (FCRN) model is then proposed to solve the winding problem on micro terminal. The model consists of three parts: (a) the Faster RCNN to detect coils from scene images; (b) a FCRN network including a full convolution network (FCN) followed by two ASPP pooling to extract the coil features, and two prediction branches, i.e. a semantic segmentation branch and a regression counting branch to predict the coil segmentation images and to predict the coil turns number, respectively; and, (c) the horizontal projection algorithm to analyse whether the prediction images represent cross coil. The experimental results show that, compared with the counting model directly using CNN classification, the FCRN counting model works well in the environment with fewer samples, and can identify cross coil at the same time of counting, which improves its adaptability and accuracy. Thus, it can improve the detection rate of products.