Image binarization method based on deep learning semantic segmentation

Abstract

The invention discloses an image binarization method based on deep learning semantic segmentation, and the method comprises the steps: 1), carrying out the real-time segmentation of a color image through a lightweight semantic segmentation network, training the network according to a BCEloss function, and obtaining a feature map after the network is converged; 2) for the feature map in the step 1), using an iterative threshold method to obtain two thresholds, using the two thresholds to segment the image into a three-valued map, and marking an area represented by an intermediate value in the three-valued map as a suspected area; and 3) for a segmentation result in the step 2), using a connected domain method to firstly denoise the image, then dividing a suspected area according to a certain rule, and finally changing the image from a ternary image to a binary image according to a division result, i.e., the corresponding foreground and background in the step 1). According to the method,the skin color detection effect is improved on the premise of meeting the real-time performance.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image binarization method based on deep learning semantic segmentation. Background technique [0002] In recent years, with the development of deep learning technology and related hardware, the term "AI" has become familiar to us. As a cutting-edge technology, deep learning has greatly improved our quality of life. Especially in the field of face, face recognition, face detection, beauty and cosmetics, deep learning technology has been widely used, bringing great social value. Skin color detection can be used as a prerequisite technology in the field of human face and human body, and can be studied as a part of the field of beauty and cosmetics and face recognition detection technology, which has high social value and significance. [0003] Traditional image processing methods almost use the color information of the image to detect skin color. Among them, one of th...

AI Technical Summary

Problems solved by technology

However, this method is the result of a large number of statistics, which has certain universality and poor robustness.

The most commonly used machine learning method is based on the mixed Gaussian model, which completes data fitting through the mixed Gaussian channel of CrCb. In practical applications, there is still the problem of false detection of color areas similar to skin color.

Images