A Strategy for Training Dim and Small Infrared Targets Detection Networks Under Sequential Cloud Background Images

The existing open-source infrared dim and small target dataset have limited data capacity and insufficient scenarios. To improve the generalization and robustness of the network, it is usually needed to establish a new dataset for the specific application scenario. When using the conventional training method in the new dataset, there are problems such as low target-to-background proportion in full-sized images, inaccurate labels due to nonexpert annotators, and inadequate diversity of data, which affect the training efficiency and network performance. This article proposes a training strategy based on the newly established dataset to overcome these problems. Specifically, small-sized image transfer learning is proposed to increase the small target proportion, shorten training time, and improve training efficiency. Moreover, a label refinement method based on loss evaluation metrics is adopted to reduce the impact of inaccurate labeling on network training. In addition, an iterative training method is proposed by supplementing new false alarm and miss detection data into the dataset between each iteration to further improve the training performance. The experiments are carried out and the results show that the above method can effectively shorten training time, improve training efficiency and performance of infrared dim and small target detection networks.