Portrait contour detection method in video based on lightweight deep neural network

Abstract

The invention discloses a method for detecting a portrait contour in a video based on a lightweight deep neural network, and the method comprises the steps: fusing cavity convolution into a MobileNetnetwork model, enabling the MobileNet network model to be lightweight, and obtaining a universal semantic segmentation model through training; designing a detail enhancement module, extracting a portrait channel through SoftMax, and combining bilinear interpolation and original image splicing to enhance portrait contour edge details; batch standardized nodes in the model are fused with front and back convolution layers, the down-sampling rate is adjusted, parameters such as the size and the depth multiplier are input to optimize the model, and the calculated amount is reduced; a WebNN API is used for realizing cross-platform performance by utilizing network application distribution, so that the method can be deployed at a client, the requirement on computer hardware configuration is reduced, and high real-time performance of video call is ensured. By applying the method of the invention, the problems of high hardware configuration requirements, incapability of cross-platform deploymentand poor real-time performance caused by rough and fuzzy portrait contour detection and large model complex calculation amount in the existing video are solved.

Description

technical field [0001] The present invention relates to the technical field of computer vision semantic segmentation, in particular to a method for detecting contours of portraits in videos based on a lightweight deep neural network. Background technique [0002] The main tasks in computer vision can be divided into: image-level image classification (image classification), object-level target detection (object detection), and pixel-level semantic segmentation (semantic segmentation). The specific meanings are respectively in the recognition diagram. What kind of objects exist, give the location and boundary of the objects in the picture, and determine what kind of objects each pixel in the picture belongs to. [0003] The fully convolutional network is the pioneering work of deep learning applied to semantic segmentation. Whether it is SegNet, U-Net, or the instance segmentation algorithm Mask R-CNN developed on the target detection algorithm Faster R-CNN, it is inseparable....

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to provide a method for detecting the profile of a person in a video based on a lightweight deep neural network, aiming at solving the problem of rough detection of a profile of a person in a video Fuzzy, complex model and large amount of calculation lead to high hardware configuration requirements, inability to deploy across platforms and poor real-time performance

Images