| Image caption is an interdisciplinary subject based on image recognition research.Its main goal is to make meaningful text caption of a given image.At present,image caption is an important research direction in the field of computer vision.Although great progress has been made in related researches,the semantic information integrity,language logic and image information loss of the relevant model proposed for the text captions of images are still not high enough.In view of these unresolved problems,this paper mainly conducts the following research:Firstly,in the aspect of image feature extraction,this paper adopts the popular deep learning method for feature expression.Different from the pre-processing method of manual extraction,deep learning USES the deep neural network for image feature recognition,analysis and extraction.On the premise of ensuring the accuracy and information integrity of the extraction process,the extraction efficiency is improved and the labor consumption is greatly reduced.Based on the comparative analysis of several existing Convolutional Neural Network(CNN)models,the InceptionV3 structure increases the Network width and reduces the computational pressure while improving the capability of feature expression.The concept of residual in ResNet structure avoids the gradient disappearance caused by redundant layers.In this paper,an improved residual-inceptionv3 model is proposed on the premise of ensuring the effect of the model.The concept of residue-inceptionv3 is integrated into the Inception layer to reduce the complexity of network computation and compete the feature extraction task of the image faster.Secondly,the feature vectors of the image are mapped to the same mapping space as the text feature dimension through the deep learning hash layer reduction,and then input into the neural network model to generate the language description text.Here using artificial Neural Network for Recurrent Neural Network(RNN)variations of structure,door control unit,this paper will be a single Neural Network structure with double differential gating cycle unit structure,designed to simulate real context analysis to a great extent,the text analyses the context semantic information in both before and after,greatly improve the accuracy of the statements generated and richness.In addition,introduced the concept of global attention mechanism,on the basis of the single-layer attention mechanism to join the "filter" module,will be part of the larger images generated probability characteristic vector information directly with the double loop neural network computing the probability that the results are integrated,not only in the very great degree of compression,the prediction of the model space at the same time also reduces the probability of missing information in text decoding model,make the probability of the larger image characteristics are more attention.This paper presents a DGIC model based on residual-InceptionV3 network and two-layer GRU network.In view of the cross-modal interaction between image and text,the deep learn-based hash layer is used to preprocess the image feature vectors before the input language model generation.In addition,the skip-grams model in Word2 Vec is selected to convert the text information in the language text library into the word embedding vector and input it into the language model.The cross-modal interaction with the image feature vector is carried out in the cyclic neural network model to generate the final text caption.In this paper,the basic data set is used as the training data set,the model is trained by the adaptive moment estimation gradient descent algorithm,and the model is tested by the cluster search algorithm.The results show that the DGIC model proposed in this paper can greatly improve the convergence rate of the model.With the deepening of training rounds,the caption text results generated by the DGIC model are more accurate,which avoids the problem of model overfitting.According to manual sampling,the text caption generated by the model is logical,more realistic and comprehensive. |
