Semanteme division method of constructing images from pixels by jointing embeddedness

Abstract

The present invention provides a semanteme division method of constructing images from pixels by jointing embeddedness. The method mainly comprises the steps of detecting the image elements, connecting pixels and jointing embeddedness and supporting the overlapping detection, and comprises the processes that a network firstly finds out all vertexes and edges of the images, each image element is based on a pixel position identified by the network, and then various images are put together; according to the steps same as the category and bounding box prediction, the network is trained to generate the additional output, for each vertex, generates an identifier in a vector embedding form, and for each edge, generates the corresponding embeddedness to call the source and the target vertex thereof and finally generates the distribution between the possible object category and the bounding box size, and the output and the available and calibrated real information are matched optimally. The supervision provided by the present invention allows the end-to-end training of the single network, the detection precision is improved, the scene image generation rate is improved substantially, and the flexibility, the generalization performance and the practicality are all enhanced.

Description

technical field [0001] The invention relates to the field of image semantic segmentation, in particular to a semantic segmentation method for constructing images from pixels through joint embedding. Background technique [0002] With the rapid development of multimedia and Internet technology, network image resources are increasing day by day, and images have become a very important information resource, which contains far more information than text, so how to fully understand the semantic content contained in images, how to truly Effectively using semantics to retrieve image resources has become an important topic nowadays. In the field of fashion, through the semantic segmentation of the human body, information such as the face, torso, and clothing can be located, thereby helping people realize functions such as automatic fitting during Internet shopping; in the field of automatic driving, through the semantic segmentation of the scene in front of the car body Segmentatio...

AI Technical Summary

Problems solved by technology

[0004] For the problem of high computational complexity, the purpose of the present invention is to provide a semantic segmentation method for constructing images from pixels through joint embedding. The network first finds all the vertices and edges that make up the image, and each image element is based on the pixel position identified by the network. Putting the various images together again, by training the network, the same steps as class and bounding box prediction produce additional output, for each vertex, the network produces an identifier in the form of a vector embedding, and for each edge, a corresponding to reference its source and target vertices, and finally the network generates a distribution between possible object classes and bounding box sizes, best matching the output to the ground-truth information available

Images