Semantic segmentation method for low-illumination scene

Abstract

The invention discloses a semantic segmentation method for a low-illumination scene, and belongs to the technical field of computer vision. According to the semantic segmentation method, a semantic segmentation problem of a normal image is used as a source domain problem; the semantic segmentation problem of the low-illumination image is used as a target domain problem; a feature migration methodin migration learning is used; and the advantage of sufficient information of a normal scene image is brought into full play, and useful information in a normal scene is extracted, and the useful information is converted and combined with feature information of a low-illumination image to obtain more image information beneficial to semantic segmentation, so that a deep neural network is trained. Based on the idea, on the basis of the generative adversarial network, a network model for direct semantic segmentation of a low-illumination scene is designed and realized by utilizing a transfer learning method. By utilizing the model, the semantic segmentation task of the low-illumination picture can be effectively solved.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to image semantic segmentation technology based on deep learning, aiming at low-light scenes, which is the most common non-ideal scene, to obtain high-precision dark scene semantic segmentation results. Background technique [0002] Semantic segmentation is a classic computer vision problem, which takes raw data such as images as input, and outputs the corresponding region-of-interest mask based on the original input. Full-pixel semantic segmentation uses a single pixel as the basic unit of classification, which is very similar to the perception of human scene understanding. Compared with the early computer vision problems that only focus on image edges and gradients, it has great advantages. Semantic segmentation can gather pixels belonging to the same part of the picture together, which can well solve the problem of scene understanding. Compared with other imag...

AI Technical Summary

Problems solved by technology

[0005] In recent years, based on deep learning technology, many semantic segmentation methods have emerged. However, due to the lack of data sets and other reasons, these methods have a single role in the scene and require ideal conditions of sufficient brightness and care. Once the brightness is insufficient, the accuracy will be severely reduced.

Images