Research On Enhancement And Extraction Of Visual Semantic Features

Visual semantics refers to the gradual accumulation of non-verbal semantic information,and the specific performance of these non-verbal semantic information in deep learning is visual semantic features.Good visual semantic features can not only improve the performance of the neural network,but also enable its internal exploration from its accumulation.Although the existing neural networks have good performance in many fields,the choice of which neural network and how to set the structure and parameters of the neural network depends largely on the experience of the researchers.Extracting the semantic features of the neural network helps to understand the internal decision logic and establish a perfect theoretical system to guide the design of the neural network.This thesis has carried out related research on the extraction of semantic features and the design of neural networks.The main work and innovations are as follows:(1)In view of the lack of readability and interpretability of the current neural network structure,a prototype system combining visual semantic features with modular neural network structure design was proposed,and three mainstream neural networks based on semantic feature constraints were designed.Improve the training speed and training accuracy of the neural network through human cognition.On the other hand,the entire neural network is easier to understand in structure,more flexible and open in design,and achieves a preliminary interpretation of the neural network.(2)Aiming at the problem of small samples and insufficient interpretation of prediction results in data prediction regression,a design method of BP prediction neural network based on known semantic feature constraints is proposed.Taking the study of geological surface reconstruction as an example,the fold structure is taken as the research object,based on human cognition,the ridge features with semantic features are first extracted,and then on the basis of the traditional BP prediction neural network,Add a semantic constraint module to adjust the backward propagation of the network,and implement a predictive neural network based on semantic constraints,so that the network can still achieve training effects with a small number of samples,and experts can make reasonable data based on semantic constraints.Explanation.(3)For the problem that it is difficult to extract language features when the semantic features in the data are unknown,a semantic feature enhancement and extraction method based on predictive neural network is proposed.Taking the geological surface reconstruction data as an example,first,the general characteristics of the form are studied,and the expert's cognition is converted into the standard form of semantic features understood by the computer through the semantic feature representation method based on the feature parameters.Based on the modular neural network design method,the adaptive neural network training process and semantic feature extraction are achieved by selecting appropriate fitting functions and secondary network training methods.It solves the problem of predicting the data when the features in the data are not clear and the semantic feature constraints cannot be extracted,and the semantic features are extracted by this method.