Real-time semantic segmentation system and method based on deep learning and weight distribution

Abstract

The invention discloses a real-time semantic segmentation system based on deep learning and weight distribution. The real-time semantic segmentation system comprises a data acquisition module, a data preprocessing module, a coding module, a decoding module, a weight distribution module and a semantic segmentation prediction module; group convolution, separable convolution in the depth direction, shuffling and the like are introduced into a coding module of the coding-decoding module, so that the calculation cost is reduced, and meanwhile, the expression capability of feature information is kept; a multi-scale fusion unit is introduced into a decoding module, feature information aggregation and an attention mechanism are used for finely processing a feature map output by a coding module, and the overall segmentation precision of a coding-decoding module is improved; in a weight distribution module, the weight of the loss value of the corresponding category is calculated by using the number of pixel points of each category in the image, so that the segmentation precision of the whole method is improved; the semantic segmentation method involved in the system is simple and easy to implement.

Description

【Technical field】 [0001] The present invention relates to the field of artificial intelligence real-time semantic segmentation, specifically a real-time semantic segmentation system and method based on deep learning and weight distribution. 【Background technique】 [0002] Semantic segmentation can be seen as a pixel-wise classification task, which can assign a specific predefined category to each pixel in an image, and this task has many potential practical applications in areas such as autonomous driving and image editing. [0003] In recent years, building deeper and larger convolutional neural networks (CNN, Convolutional Neural Networks) is the main trend in solving semantic segmentation tasks. Most CNN networks that only aim for accuracy usually use hundreds or thousands of feature channels and convolutional layers. Although higher accuracy has been achieved, in many realistic application scenarios, such as augmented reality, robotics, and autonomous driving, etc., sma...

AI Technical Summary

Problems solved by technology

For example, the Lightweight Encoder-Decoder Network (lightweight codec network) uses Split-Shuffle-non-bottleneck (segmentation-shuffling-non-bottleneck unit) as a residual layer to perform effective inference calculations, which uses the classic The codec structure greatly reduces the network parameters, however, the decoding module cannot flexibly use the features of different layers to maintain better accuracy, the number of network layers is small, and it cannot extract enough feature information to achieve accurate classification

Images