Research And Implementation Of Image Title Generation Technology Based On Deep Learning

In recent years,with the popularity of the image acquisition device in various industries,and it is closely combined with the actual life in the field of application of generated a lot of digital image,image title generation is more and more widely used.In real life,infants and young children early education,visual loss or visually impaired all could benefit from the image header generation algorithm.For example,the semantic information contained in the image is effectively described in words,and read out in the way of voice broadcast to help the blind to understand the surrounding environment in advance,so as to prevent dangerous situations.Therefore,how to make the computer better understand the content of the picture like people,and organize language to express the content of the picture like people,which has important research significance.In this paper,the key technologies of object detection in natural scenes,title generation from image to text and title generation from specific character images are studied.The main research work is as follows:1)A target detection method based on fast convolutional neural network is proposed.The image is continuously convolved and pooled based on faster and deeper volume network,and the image feature image output by the convolution layer is shared by the subsequent regional suggestion network and the full connection layer,so as to quickly extract the high-dimensional abstract features of the image.The fusion region suggestion network is used to extract the regions where there may be entities in the image,and finally to identify the regions containing objects in the image.2)An image title generation method based on FR-LSTM-ATT natural scene is proposed.The open source image title data set is used to construct the language model based on attention mechanism recurrent neural network.The convolution devolution network and regional suggestion network in target detection method are used to obtain the entities existing in the target image,and the two are fused to complete the image title generation task.3)This paper proposes a Captcha-IC based title generation method for specific character images.From the perspective of image title generation algorithm in natural scene,the problem of converting a specific character image into a text title is solved,and the network structure is adjusted according to the characteristics of a specific character image.The idea of image title generation algorithm is applied to the conversion of the specific character image to the text on the premise of following the code-decoding structure,eliminating manual operation.By optimizing the network structure,the method of detecting the character content in the specific character image and combining with sentence template matching is realized.The text content of the detected image is fused into the sentence template to generate a sentence as the title of the character image.Experimental results show that the proposed method is effective.The method in this paper is effective.The proposed target detection model is aimed at the target detection models learned from Pascal Voc2007 and MSCOCO2014,which are two public datasets,and the m AP value of 20 target attributes evaluated on Pascal Voc2007 test set reaches 70.1%and 72.5% respectively.It can effectively realize the positioning and recognition of the target in the scene image of daily life.In the image title generation model based on FR-LSTM-ATT,the evaluation indexes BLEU?1 and BLEU?4 reach 74% and 32%respectively.It can effectively transform natural scene image into text.Captcha-IC based on the specific character image title generation algorithm can effectively express the character image in a sentence.