Research On Similarity Calculation Method Of Depth Perception Image Based On GAN And Feature Fusion

Comparing the differences between different sample data is the basis for many applications and research efforts.In this thesis,the similarity calculation based on convolutional neural network model and depth perception image is studied.Firstly,the depth perception measurement method based on convolutional network is proposed for the problem of inaccurate and poor effect of traditional Euclidean distance measurement.Secondly,it is based on traditional methods and networks.The problem of single feature and lack of semantic information is extracted,and the feature fusion model of fine-tuning VGG is proposed.Finally,the Atrous-MobileNet model is proposed for the problem of large amount of calculation and multi-parameters for perceptual metric and feature fusion model,and applied to image generation,respectively.Image retrieval and image recognition tasks are verified.The innovations of this thesis mainly have the following three aspects:First of all,an image latent feature learning method based on GAN and depth perception metrics is proposed.Firstly,the feature map extracted by the VGG network is used as the training loss,so that the features extracted by the network are more semantically similar.Second,the maximum mean difference metric is used to map the image to the regenerated Hilbert space to measure image differences.By comparing and analyzing the experimental results of different conditions,it is proved that the proposed model has the characteristics of improving potential feature learning and feature similarity measurement.In addition,the influence of pixel loss and depth perception metric on image generation in training dataset is also discussed.Secondly,an image retrieval method based on multi-feature fusion and model fine-tuning is proposed.Aiming at the problem that the traditional image feature extraction method has a single feature and weak semantic information,this thesis firstly fixes the shallow features of the VGG network model(first three layers)and loads the pre-training model to fine-tune the network.Then,add multiple levels.The feature fusion of the feature pyramid structure realizes the fusion of multi-channel multi-scale convolution features with the underlying details and deep semantic features.The experimental results show that compared with the previous model,the method proposed in this thesis has better performance and higher accuracy in image retrieval tasks.In addition,this thesis combines the PCA and hash mapping methods in the retrieval process to further improve the performance of image retrieval tasks.At last,a face detection and recognition algorithm based on Atrous-MobileNet and face key points is proposed.Firstly,for the problem that the sensory metric and feature fusion are large in calculation and many parameters in the recognition task,the Atrous-MobileNet model is proposed for the classification and recognition of glasses.Secondly,in order to deal with the scale and rotation of the image,the image is regionally normalized.The area of the eye is detected according to the position of the face of the face.Through a large number of experiments,the glasses recognition model proposed in this thesis has achieved good results in the three classifications of glasses(with,without,sunglasses)and glasses,and the improvement in robustness and recognition accuracy,and the calculation of the model.And the parameters are reduced.