Image Description Method Research Based On Convolution Recurrent Network Integration Training

The image description task is to generate a corresponding textual description of the input image,which is conducive to a better understanding of different visual scenes.This task has attracted much attention in the field of artificial intelligence,and it is of great significance in terms of driverlessness and military security.This paper studies the methods used in the current image description tasks,and proposes an integrated training method based on the existing methods to improve the effect of image description.The main research work and innovations are as follows:1.Complete image preprocessing and image feature extraction for the image description data set.In terms of preprocessing,this paper screens all the image data in the data set,and deletes and transforms the damaged image,grayscale image,and multi-channel image that do not meet the requirements in the original image.In the aspect of image feature extraction,the convolutional neural network is used in this paper,and the migration learning method is used to migrate the pre-trained parameters to the convolutional neural network in this paper,and the network parameters are adjusted according to the existing data set to improve The speed of feature extraction improves the effect of feature extraction.2.The text preprocessing work is completed on the image description data set.In this paper,the text preprocessing operation is mainly to divide the text into words and extract the keywords,and select the words with higher frequency for vector conversion.This article uses a long-term memory network to train based on the relationship between images and text.3.This paper introduces the existing loose model of separate training,and based on the model,proposes a joint model of combined training of convolutional neural network and long-short memory network.Through the introduction and analysis of the loose model and the joint model,the advantage of the joint model is explained,and the core integrated training method is introduced in detail.In order to maintain structural integration,the TensorFlow framework was chosen to complete the construction of the joint model.In the forward propagation process,the features of the image are extracted using the convolutional neural network,and the extracted image features are used to initialize the hidden layer of the long-term and short-term memory network.Meanwhile,the text vector is used as the input of the long-term and shortterm memory network,thereby making the image Associated with text.In the back propagation process of the training,the parameters of the two networks are updated simultaneously according to the final loss value,and the integrated training is completed,and the image is implemented as an input text description as an end-to-end generation method of the output.This paper chooses to use the loose model of separate training and the joint model of integrated training for experimental comparison.The experimental results show that the integrated training method proposed in this paper realizes the end-to-end method to simplify the training operation process,and through the data analysis of the BLEU score,the score of the joint model is improved by 18.08% compared with the loose model score.The visual score distribution is significantly better than the loose model.The comparison results show that the image description effect of the joint model of integrated training proposed in this paper is better than the loose model of separate training.