It is a challenging task to separate infrared small targets from complex backgrounds quickly and accurately. With the development of deep learning, to effectively extract the features of infrared small targets, many algorithms use classic networks (VGG, UNet, or ResNet, etc.) as a backbone in the encoding phase. At the same time, various multiscale feature fusion modules, channel, and space attention mechanism fusion modules are designed. Although these designs are slightly helpful to the improvement of infrared small target detection accuracy, they will cause a significant increase in params and floating-point operations per second (FLOPs). To reduce the computational complexity, a lightweight infrared small target segmentation network (LW-IRSTNet) is proposed. The Depthwise separable-Atrous-Asymmetric-Atrous (DAAA) module is built by taking full advantage of the combination of depthwise separable, atrous and asymmetric convolution. The ablation experiment shows that the DAAA module plays an important role in improving segmentation accuracy and reducing params and FLOPs. To verify the accuracy, robustness, and computational complexity of LW-IRSTNet, 17 state-of-the-art networks are used as baselines for comparative analysis. The experimental results showed that the segmentation accuracy indexes (mIOU, F1, ROC) of LW-IRSTNet are all above or equal to the baseline best results on the five public datasets. Meanwhile, the network params are compressed to 0.16M and FLOPs to 303M, which is much lower than the baseline results. In addition, LW-IRSTNet is deployed on the Orange Pi 5 embedded platform. Through the ONNX framework, NPU acceleration and CPU multi-threaded resource utilization, the high-performance inference capability of LW-IRSTNet is realized, with FPS up to 50. A lot of comparative experimental results and the LW-ISTSNet inference model based on the Ptorch and onnx frameworks is published at https://github.com/kourenke/LW-IRSTNet.