The Application of Convolution Neural Network in WheelHub Classification

In this work, through investigating traditional methods about abnormity detection of wheelhub, we find that it takes a lot of time to extract specific features and set specific threshold parameters for the specific classification of wheelhub. In order to looking for an easy way to solve the general classification of wheelhub task, a very deep convolution network of 16 layers with the pre-trained parameters which is named PreVGG-16 is proposed. To prove the reliability of our network, firstly we adjust the network structure of very deep convolution network with 16 layers (VGG-16) properly. Then we use pre-trained parameters and randomly initialized parameters to train the network respectively. With the experiment, we have demonstrated that the network of PreVGG-16 which has trained a large number of images is far ahead of the randomly initialized parameters of network. With the limited training dataset, our experiment has achieved 6.00% top-1 error on the test set of the wheelhub images in PreVGG-16 network without specific feature extraction and threshold setting. As a result, we present a new idea to solve the problem of general wheelhub classification and achieve a reliable result.