An Image Representation Study Of Dream Recall

In recent years,with the widespread application of deep learning methods in the fields of natural language processing and computer vision,more and more researchers have focused on the intersection of language and vision.Specifically,the task from text description to image generation refers to the generation of images that conform to the semantics of the text through the generation network according to the given text information.Although,the research of generating images from simple categories of text descriptions such as flowers and birds has made exciting progress.However,there are still no satisfactory results in the research of generating images from more complex text descriptions.The main problem to be solved in this article is how to generate images based on the textual description of complex dreams.From a macro perspective,dreams are the inspiration of some major creativity and scientific discoveries in human history.Recording dreams in the form of images is an interesting and meaningful thing.From an application point of view,dreams are scenes that often appear in novels,articles,and comics.If images can be automatically generated from the description of dreams,it will greatly reduce the cost of text creators and improve the creative efficiency of cartoonists.Compared with the conventional text generation image task,first of all,the dream scene contains a large number of objects and complex relationships.How to accurately extract the objects and relationships is an important issue.In addition,scenes in dreams are more fantasy than real scenes.How to predict the layout of scenes in dreams is another important issue.Finally,the dream image and the psychological emotions reflected behind it are also an important factor in visualizing dreams.In response to the above problems and in order to generate more realistic dream images,this article mainly conducts research from the following three aspects:(1)Object-relation extraction algorithm based on dream text.This new dream-based object-relation extraction algorithm proposed in this paper effectively solves the problem of object confusion in information extraction.In general text information extraction,the objects in the text and the relationship between the objects and the objects are extracted.However,when there are two or more objects of the same category in the text,the information extracted by the traditional algorithm text will cause confusion.Therefore,this algorithm adds a de-duplication mechanism,adds objects to the collection,uses the index of the objects in the collection to refer to the objects,and distinguishes different objects according to the input of text information,so as to prevent the confusion of the object-relationship.(2)A dream scene graph synthesis algorithm based on graph convolutional neural network.Generally,the generation of images is trained in the confrontation generation network,but for the speciality and complexity of the dream scene,a single confrontation generation network cannot be solved.To solve this problem,this paper proposes a multi-stage generation algorithm centered on the scene graph.Different from the method of directly generating images in traditional algorithms,this algorithm first uses word segmentation to extract the information in the text to form scene graph elements,and then uses graph convolutional neural network to predict the scene layout,thereby solving the phenomenon of strange positional relationships in dreams.This algorithm uses a graph structure to represent the objects in the dream image and the relationship between the objects and the objects,and uses the predicted bounding box and segmentation mask to integrate the layout of all objects.(3)Dream mood rendering algorithm based on cyclic consistency network.This article adds the influence of dream emotions on the basis of image generation.The emotions in the dream have a great influence on the dreamer,and different dream emotions will produce different dream interpretations.Based on this,this paper proposes the first dream emotion data set,which contains five categories:calm,anxiety,sadness,grotesque and happy,with 1000 data in each category.Different traditional algorithms require paired image data and the situation with a single style.This algorithm is based on the Cycle-Consistent Adversarial Networks(CycleGAN),and uses the method of image style conversion to train non-paired image data to convert different emotions and generated pictures,Complete the rendering of dream emotions.