Infrared Image Semantic Segmentation Based on Improved DeepLab and Residual Network

Abstract

In the infrared temperature measurement system for non-contact online temperature detection, we establish a mapping model between grayscale image and temperature variable of electrolyte based on the principle of infrared thermography. In order to eliminate the interference of the floating material and impurities on the electrolyte image, it is necessary to accurately divide the electrolyte in the image. Therefore, this paper uses the deep learning method to construct the framework of the semantic segmentation of the aluminum electrolyte image, that is, the DeepLab framework based on the residual network ResN et-l 0 1 convolutional neural network, which is formed by the cascade of the mature modules of ResNet and improved CRFs, solving the problem of network degradation caused by network deepening. The basic structure of this network can include the following parts: First, the DeepLab framework adopts data augmentation transformation to prevent over-fitting of the network. Secondly, this framework removes the loss of semantic information. A large pooling layer uses hole convolution to calculate feature maps with higher sampling density. In addition, the ASPP (atrous spatial pyramiding pool) module performs parallel sampling with a hole convolution at different sampling rates on a given input, which is equivalent to capturing the context of the image in multiple ratios and improving the resolution of feature extraction. Finally, the improved CRF-RNN which is combined with context image information is used in the segmented processing link to smooth the noise segmentation diagram and enhance the ability of the model to capture details. The frame model of this paper can meet the requirements of image segmentation in industrial temperature measurement.