Workpiece Detection Based on Image Processing and Convolutional Neural Network

This paper presents a method on workpiece detection based on image processing and convolutional neural network(CNN). Firstly, four extreme points and center point of the workpiece are detected by image processing technologies such as canny edge detection operator, morphological processing and denoising processing. And the predicted boxes fitting the shape of image is generated. Then, according to the coordinates of the extreme points, the image with a single workpiece is cut out, and a novel CNN named workpiece-net(wp-net) is created to classify the object. As a result, the accuracy of image cutting is 0.9986, the average Intersection over Union(IoU) is 0.9235; the parameter size of wp-net is 98.25K and the average precision of classification is 0.9883. In addition, the average recall of classification is 0.9877 and the speed of classification is 0.1243s/fps without Graphics Processing Unit(GPU) and multithreading. Compared with the pure deep learning method, this method can detect more accurate coordinates which are consisted of extreme points. At the same time, the number of wp-net parameters and the complexity of the model structure used for classification are so far less than the popular deep neural network for detection that it can be easily deployed in embedded devices with limited storage space and computing power.