Research On Object Detection And Style Transfer Algorithm Based On Deep Learning

With the rise of artificial intelligence technology,especially the rapid development of deep learning and convolutional neural network,a large number of advanced theories and algorithms have emerged in the field of computer vision.Object detection and style transfer as the important research directions in computer vision,which becomes hotspots for many researchers.In current vision algorithms,object detection is undoubtedly the most important technical link,and the accuracy of object detection has a great impact on subsequent object recognition(such as a face).In order to increase the accuracy of general object detection,this paper proposes an improved R-FCN algorithm called Accurate R-FCN which based on the relation module.By introducing the relation module to make an interaction between their appearance and geometric features,the detection and recognition accuracy can be improved.In addition,to solve the problem of position deviation in the region of interest(Ro I)quantization,a position-sensitive region of interest queue(PSRo I align)is applied with the Region Proposal Network(RPN),which generates the accurate region proposal.Aiming at the specific object of face,this paper proposes an improved R-FCN face detection algorithm called Accurate Face R-FCN which based on Position-sensitive average pooling(PS average pooling).PS average pooling is used to give different weight to various areas of the face to complete accurate face detection.On the basis of accurate detection of face objects,in order to achieve the style transfer of face images,this paper improves the network structure based on the Cycle-consistent Generative Adversarial Network(Cycle GAN),eliminating the cycle network which affects the training convergence speed.During the image generation phase,the prior information of the object domain and the original image are cascaded in depth;the loss function is optimized by using the classification loss instead of the cycle consistency loss,so that the improved Cycle-consistent Generative Adversarial Network(Cycle GAN++)can be used to realize style transfer on the face images extracted by Accurate Face R-FCN in a specified style.In order to evaluate the performance of the algorithm,three algorithms which based on general object detection,face detection,style transfer proposed in this paper were experimentally verified and evaluated on the three public test datasets of COCO,Wider Face and Celeb A.The experimental results show that the Accurate R-FCNalgorithm achieves higher detection accuracy in the COCO test set than the other seven classic algorithms;the Accurate Face R-FCN algorithm compares with the twelve other face detection algorithm in the Wider Face test set which achieved better results;in the Celeb A test set,Cycle GAN++ performed better than the other four style transfer algorithms.