Network model capable of jointly realizing semantic segmentation and depth-of-field estimation and training method

Abstract

The invention discloses a network model capable of jointly realizing semantic segmentation and depth-of-field estimation, and the network model comprises a feature sharing module and a multi-task sub-network; the multi-task sub-network comprises a plurality of task sub-networks with the same structure for processing different task targets, and comprises a feature screening module, an attention concentration module and a prediction module; the feature screening module screens out features related to the task from the shared features; the attention concentration module is used for improving thecorrelation between the screening features and the task target; and the prediction module is configured to output a processing result of each task target after convolution of the concentrated attention features. The invention further discloses a training method of the model. Back propagation iterative training is carried out on semantic segmentation and depth-of-field estimation. The model provided by the invention is high in accuracy, strong in robustness and light in weight.

Description

technical field [0001] The invention relates to computer vision image processing, in particular to a network model and a training method that can jointly realize semantic segmentation and field depth estimation. Background technique [0002] Semantic segmentation is a typical computer vision problem. It belongs to high-level visual tasks and is an effective way to understand scenes. In a micro sense, semantic segmentation predicts all pixels in the image and marks each pixel with the category label. It is also an important step in achieving fine-grained reasoning. For fine-grained reasoning, the positioning and detection of objects will not only require object category information, but also require additional information about the spatial position of each category, such as center points or borders, so semantic segmentation is an important step in realizing fine-grained reasoning . [0003] In the existing application scenarios of computer vision image processing, such as ...

AI Technical Summary

Problems solved by technology

[0005] Purpose of the invention: The purpose of this application is to provide a network model and training method that can jointly realize semantic segmentation and depth estimation, which is used to solve the problem that the single-task model cannot perform semantic segmentation and depth estimation at the same time in the prior art. The defects of unsatisfactory force concentration effect and large amount of calculation

Images