A multi-resolution feature extraction convolution neural network is proposed for the problem of inaccurate edge segmentation when existing image semantic segmentation algorithms process low-resolution infrared images. DeepLabv3+ is used as the baseline network and adds a multi-resolution block, which contains both high and low resolution branches, to further aggregate the features in infrared images. In the low-resolution branch, a GPU friendly attention module is used to capture high-level global context information, and a multi-axis-gated multilayer perceptron module is added in this branch to extract the local and global information of infrared images in parallel. In the high resolution branch, the cross-attention module is used to propagate the global features learned on the low resolution branch to the high resolution branch, hence the high resolution branch can obtain stronger semantic information. The experimental results indicate that the segmentation accuracy of the algorithm on the dataset DNDS is better than that of the existing semantic segmentation algorithm, demonstrating the superiority of the proposed method.