Image semantic segmentation method based on multi-level feature fusion and Gaussian conditional random field

Abstract

The invention discloses an image semantic segmentation method based on multi-level feature fusion and Gaussian conditional random field, including: 1) constructing an image pyramid; 2) using atrous convolution to keep the resolution of the feature map unchanged; 3) multi-level features Layer fusion tuning architecture; 4) upsampling with bilinear interpolation; 5) definition of loss function; 6) optimized output of Gaussian conditional random field. The present invention realizes the full convolution architecture of layer-by-layer fusion of multi-level features by constructing an image pyramid, and uses a top-down optimization framework to replace the previously popular parallel pooling module. The features between the adjacent layers of the pyramid are preferentially fused to capture the context information to the maximum extent, and the Gaussian conditional random field is used to further optimize the front-end output, capture more spatial details, and make the object boundary in the segmentation effect map more accurate, and finally the output of the overall architecture Get the best semantic segmentation effect.

Description

technical field [0001] The invention relates to an image semantic segmentation method based on multi-level feature fusion and Gaussian conditional random field. Background technique [0002] Early image segmentation only roughly divided the content of the image into several regions, but with the development of research, the roughness of image segmentation can no longer meet the needs of various applications, and then semantic segmentation was proposed. The semantics of an image refers to semantic information such as the category of objects or entities contained in an image or an image region, and image segmentation under semantics is called semantic segmentation. Image semantic segmentation can separate the foreground from the background in a single frame image, and identify the category of each foreground object, which is equivalent to assigning a semantic label to each pixel. Semantic segmentation is a major upgrade of image segmentation in both accuracy and fineness. ...

AI Technical Summary

Problems solved by technology

The standard convolutional neural network architecture is not good at dealing with the balance between local information and global information. The pooling layer can make the network achieve a certain degree of spatial invariance and maintain the same computational efficiency, but loses the global context information. The convolutional network without pooling layer is also limited, after all, the receptive field of its neurons can only grow linearly with the number of layers

[0009] 2. The optimization of the back-end model is difficult to converge: the back-end introduces a conditional random field to optimize the output of the segmentation architecture to enhance its ability to capture details

However, whether it is the commonly used Gibbs conditional random field or Markov random field, there are problems such as complex formulas and invariant solutions. These problems often lead to slow convergence, difficult convergence or even non-convergence during model training.

Images