Improved Residual Learning Model For Hardware Image Classification

Hardware is a common product in handicraft production.With the rapid development of digital economy,manufacturers need to informatization of hardware.An accurate image recognition system can improve the efficiency of manufacturers in completing product informatization.An image classification model with high accuracy and strong generalization ability is the key to build a good image recognition system.This thesis mainly solves the problem of class imbalance of image dataset(hardware image)in image classification,and proposes an image augmentation method based on improved CycleGAN.At the same time,an improved residual network model is proposed to solve the problem that ResNet cannot extract the most efficient features in the image recognition technology of hardware,which leads to poor classification effect.The main research contents of this thesis are as follows:(1)There is a serious class imbalance problem in the original hardware image dataset.To solve this problem,an image augmentation method based on CycleGAN is proposed to enhance the dataset by generating images with few sample classes.At the same time,for the problem of image quality generated by CycleGAN are not good,we improved CycleGAN,that some traditional convolution layers in CycleGAN are replaced with Depthwise Over-parameterized Convolutional Layer(DO-Conv).In order to verify the improvement effect,a comparative experiment is carried out on the hardware image dataset.The experiment shows that the improved CycleGAN model is better than the original model in terms of image quality and diversity.After the image augmentation of the dataset,the effect of the classification model has also been greatly improved.(2)For the problems of large intra class gap and small inter class gap in the image dataset of hardware,an image classification method based on improved residual network is proposed,which improves the model by adding DO-Conv and Pyramidal Convolution(PyConv)to ResNet.The improved model can extract different levels of features,and the amount of parameter that can be learned is also higher than the residual model before the improvement.In order to verify the effectiveness of the improved model,the thesis conducted a number of comparative experiments.First,experiments are carried out on the hardware dataset and the fine-grained image dataset.The experiments show that the improved model is effective.Secondly,the Ablation Experiment of the improved residual model is carried out on the hardware dataset.The experimental results show that the image classification effect of the improved residual model is better than that of the partially improved model using PyConv and DO-Conv alone.Finally,on the hardware image dataset and the dataset expanded by CycleGAN and improved CycleGAN,the improved residual model is used for comparative experiments.On the dataset expanded by improved CycleGAN,the classification effect of the model is the best,with an accuracy of 95.87%,a recall of 91.82% and an F1 score of 93.58%.The effectiveness of the image augmentation method based on improved CycleGAN is further verified.Experiments show that the image data enhancement method based on improved CycleGAN proposed in this thesis can effectively alleviate the class imbalance of hardware image dataset,and the classification effect of improved residual model in hardware image is also good.