Essential image analysis method based on multi-scale attention and label loss

Abstract

The invention provides an essential image analysis method based on multi-scale attention and label loss. A cyclic convolution attention mechanism and an adversarial idea are introduced into an essential decomposition problem, a multi-scale attention MSA-Net network used for essential image analysis is constructed, and a network structure follows a basic framework of a generative adversarial network (GAN) and comprises a generator and a discriminator. The generator is composed of an attention sub-network and a codec sub-network, and is used for decomposing an image into a reflection image and an illumination image. The discriminator is used for giving the probability that any input image is a correct essential image. Meanwhile, the invention also provides a new label loss function for improving the decomposition effect of the reflection graph, and the loss function is constructed based on a label image in a data set, so that the reflection graph obtained through network decomposition has a better local texture consistency effect and a better quantitative evaluation index.

Description

technical field [0001] The invention belongs to the field of image processing, and in particular relates to an essential image analysis method. Background technique [0002] Image understanding and analysis is one of the important basic research in the field of computer vision. In complex natural scenes, the same target may have differences in image surface color, grayscale mutations, etc. due to many factors such as light intensity, shadow occlusion, and attitude changes, resulting in huge differences in the observation effect of the same object in the same scene. If the image is directly processed, it will greatly increase the difficulty of image analysis and understanding, which will affect the performance of the algorithm. To solve this problem, the best way to deal with it is to dig out the inherent mode of the target object in the image—essential features, and then send the essential features of the object to the subsequent algorithm for processing. Essential feature...

AI Technical Summary

Problems solved by technology

However, since the features in the reflection map and the light map do not satisfy the completely mutually exclusive characteristics, the network often cannot achieve the ideal separation effect, and the separation of the reflection map and the light map is not complete, so a more ingenious network structure and loss are required. Function design to further improve the quality of essential image analysis

Images