Research On Object Detection And Image Segmentation Based On Deep Convolutional Networks And Spectral Representation

With the improvement of computer performance and the emergence of large image datasets,deep learning has set off a new wave in the field of image processing and computer vision.Due to the good feature extraction ability of Convolution Neural Network(Convolutional Neural Network,CNN),CNN is widely used in object detection and image segmentation,the two basic tasks of computer vision.There emerges many feature extraction and fusion methods such as residual module,feature pyramid,dilated convolution,skip connection and attention mechanism.The spectral representation is an important method of traditional image processing,which has a strong ability of feature representation and provides a unique perspective for image feature extraction.Therefore,aiming at the improvement of feature extraction ability and more effective information fusion method of deep convolutional network,this paper proposes to introduce spectral representation into deep learning model to study the object detection and image segmentation network model.The specific work are as follows:1.This thesis proposes a medical image segmentation method based on deep convolutional network and spectral representation.The spectral representation is introduced into the deep convolutional network model to construct a new convolutional layer.The advantages of Fourier transform and its inverse transform which combined spatial domain and frequency domain are utilized to improve the effect of feature extraction.In the frequency domain,the high frequency and low frequency of input features are separated to construct a new semantic information fusion method.The extracted shallow features and deep features are fused based on spectral representation,and better medical image segmentation results are achieved.Evaluated on three datasets—skin lesion segmentation(ISIC 2018),retinal blood vessel segmentation(DRIVE),lung nodule segmentation(LUNA),and brain tumor segmentation(Bra TS2019),the proposed model achieves outstanding results: the metric F1-Score is 0.878 for ISIC 2018,0.8185 for DRIVE,0.9909 for LUNA,and 0.8457 for Bra TS 2019.Experimental results show that the proposed method can enhance the feature extraction ability of the convolutional network and achieve more accurate segmentation results.2.This thesis proposes a deep convolutional network model for object detection and image segmentation based on spectral representation and feature pyramid.On the basis of the one-stage network model which can simultaneously accomplish object detection and instance segmentation,the feature pyramid is introduced to fuse multi-scale information to solve the problem of object resolution change.The spectral representation is introduced to improve the capability of feature extraction.The proposed model obtains the position coordinates,category information and pixel segmentation results of the target object simultaneously.Evaluated on PASCAL VOC2012 dataset,the MAP values of object detection and instance segmentation were75.8% and 72.9%,respectively.Experimental results show that the proposed model achieves better detection and segmentation performance compared with other models.