Parallel multi-scale attention mechanism semantic segmentation method and device based on deep learning

Abstract

The invention discloses a parallel multi-scale attention mechanism semantic segmentation method based on deep learning, and the method comprises the steps: firstly carrying out the preprocessing of animage data set, and improving the segmentation precision and robustness of a model; based on ResNet50 obtained after adjustment of a fifth convolution layer, aggregating multi-scale semantic information through a parallel multi-scale attention module connected to the top of a base network. And finally, recovering the image size through bilinear up-sampling. According to the method, a similarity EDS module is added behind a feature map obtained by five original parallel convolution kernels with different sizes. By adding the attention mechanism, important semantic information in a feature mapobtained by five parallel expansion convolution is enhanced, and secondary semantic information is inhibited.

Description

technical field [0001] The invention belongs to the field of deep learning and computer vision, and in particular relates to a deep learning-based parallel multi-scale attention mechanism semantic segmentation method and device. Background technique [0002] Semantic segmentation is a basic and challenging task. Its purpose is to predict the category of each pixel, that is, to learn high-level semantic information and local location information of object outline, object position and object category. As one of the most basic tasks of computer vision, semantic segmentation has been widely used in areas such as autonomous driving, medical diagnosis, video editing, object detection, and aerial image analysis. In recent years, with the development of deep convolutional neural networks, compared with traditional machine learning methods such as random forests, deep convolutional neural networks have more powerful feature extraction capabilities. In particular, the emergence of th...

AI Technical Summary

Problems solved by technology

Although ASPP can effectively capture multi-scale information by using several convolution kernels with different hole rates, it does not take into account the differences between the features captured by different receptive fields in the multi-scale information aggregation stage.

Images