A lightweight real-time infrared image instance segmentation algorithm model is proposed. Experiments are carried out on the infrared image data set. Aiming at solving the problem that the segmentation accuracy of the model is not high, the following two improvements are made: firstly, the improved tiny Yolo backbone network is added with feature pyramid network to refine the features, highlighting the boundary and small object features; Then, aiming at the problem that large objects account for more in the loss function, three different sizes of anchor, prediction head and prototype mask networks are introduced, which are connected to the segmentation model for end-to-end training, so that small objects can make greater contributions to the network. Experiments show that the lightweight real-time infrared image segmentation model achieves higher segmentation accuracy for small objects.
|