Parts Recognition Based On Convolutional Neural Networks

Parts recognition and classification task is a very common task in industrial production process.It can classify different kinds of target parts,or sort out parts with quality defects.Nowadays,most of the parts recognition algorithms are basically implemented by extracting a series of features from the image of the parts,and then feeding these features to a na?ve machine learning algorithm.However,extracting features manually not only needs a huge amount of work,but the selected features a usually not optimal.Convolutional neural network can realize extraction features from the image automatically and connect it with classification process naturally,which can achieve self-learning through data.Therefore,this paper raises a pattern recognition algorithm based on convolutional neural network and realized parts recognition and classification task.Firstly,the image data of the parts being classified is acquired,and data augmentation is implemented.Then,a simple convolutional neural network structure is designed by combining the structure and the design method of convolutional neural network with the specific characteristics of the images being classified and the demand of the problem.This convolutional neural network consists of 4 convolutional layer,3 max pooling layer and 2 fully connected layer,which achieves basic parts classification task.When using the network structure mentioned above classifying parts,as some parts with similar shapes and sizes are difficult to classify,which leads to low accuracy rate,an image preprocessing algorithm based on edge detection and maximum connected domain called target-centralization algorithm is raised in this paper.It can extract the target area in the image and move it to the center of the image.The principle and function of target centralization are expounded theoretically,and the calculation process of target centralization algorithm is described concretely.The experiment results shows that the accuracy of the model can be improved significantly by implementing the target-centralization algorithm.In order to solve the malpractice that convolutional neural network has a large amount of parameters and calculation,which is difficult to apply to mobile devices,the structure of the net is improved by replacing fully connected layers by global average pooling layer and 1×1 convolutional layer,which reduces a large amount of parameters.The experiment results shows that the accuracy rate of the optimized structure doesn't drop basically,which means the model gets a good trade-off between accuracy rate and number of parameters.The results of this paper prove that convolutional neural network can classify different shapes and sizes of parts at the same time.It not only can avoids the complex artificial design processes of selecting features and building template library,but have better performance and accuracy than traditional machine learning algorithm as well.The target-centralization algorithm raised in this paper can improve the accuracy among parts with similar shapes and sizes.It is suitable for images that have clean background and small target.