Hybrid Dilated Convolution Neural Network For Image Classification And Object Detection

In the field of computer vision and pattern recognition,the convolution neural network(CNN)is a widely used technology,in recent years,with the development of deep learning theory,traditional CNN model cannot meet the increasing demand very well due to its lower performance and calculation of excessive resource consumption,therefore,new methods that can improve the efficiency of image classification and object detection are needed.In this paper,the dilated convolution model and HDC model for image classification are studied,and the HDF-RCNN model for image object detection is proposed.The main works are as follows:1)the relevant basis of CNN and object detection are deeply studied,the principle and structure of dilated convolution are analyzed,a dilated convolution model is built,and the good performance of dilated convolution model in image feature extraction on Mnist handwritten digital recognition data set is verified.The experimental results show that the training time of dilated convolution model is 11.31% lower on average and the training accuracy is 0.5% higher than that of traditional CNN,but its testing accuracy is 0.18% lower;2)in view of the shortcomings in dilated convolution model,an HDC model is designed and built to form a complete convolution kernel without cavities by stacking dilated convolution kernels of different sizes.Such a convolution kernel is used to process the input images.The performance of HDC model,the traditional CNN and the dilated convolution model with the same structure are tested and compared on the wide-band remote sensing image data set,which verifies the superiority of HDC model.The experimental results show that the training time of HDC model is 2.02% lower than that of dilated convolution model,and the training accuracy and testing accuracy are 14.15% and 15.35% higher,respectively;3)the HDF-RCNN(hybrid dilated Faster RCNN)model is designed and built,and the LeakyReLU activation function is used to replace the ReLU activation function,so as to further reduce the impact of dilated convolution.Microsoft COCO data set is used to test the HDFRCNN model and the image object detection results of HDF-RCNN model are obtained.The experimental results show that compared with the traditional Faster RCNN,the training time of HDF-RCNN model is 34.29% lower on average,the training and testing accuracy are 40.06% and 7.11% higher,respectively,and the average recognition rate of 1200 images in 10 categories is 0.6% higher.The HDF-RCNN model proposed in this paper solves the problem of information loss in dilated convolution,and achieves good results on Mnist handwritten digital recognition data set,wide-band remote sensing image data set and Microsoft COCO data set.It is a method of image object detection with good performance.