红外目标检测网络改进半监督迁移学习方法

针对红外数据集规模小，标记样本少的特点，提出了一种红外目标检测网络的半监督迁移学习方法，主要用于提高目标检测网络在小样本红外数据集上的训练效率和泛化能力，提高深度学习模型在训练样本较少的红外目标检测等场景当中的适应性。文中首先阐述了在标注样本较少时无标注样本对提高模型泛化能力、抑制过拟合方面的作用。然后提出了红外目标检测网络的半监督迁移学习流程:在大量的RGB图像数据集中训练预训练模型，后使用少量的有标注红外图像和无标注红外图像对网络进行半监督学习调优。另外，文中提出了一种特征相似度加权的伪监督损失函数，使用同一批次样本的预测结果相互作为标注，以充分利用无标注图像内相似目标的特征分布信息；为降低半监督训练的计算量，在伪监督损失函数的计算中，各目标仅将其特征向量邻域范围内的预测目标作为伪标注。实验结果表明，文中方法所训练的目标检测网络的测试准确率高于监督迁移学习所获得的网络，其在Faster R-CNN上实现了1.1%的提升，而在YOLO-v3上实现了4.8%的显著提升，验证了所提出方法的有效性。

An improved semi-supervised transfer learning method for infrared object detection neural network

In view of the infrared datasets which has limited scale and few labeled samples,a semi-supervised transfer learning method was proposed for the training of infrared object detection neural network.It aimed at improving the training efficiency and generalization ability of object detection neural networks on infrared datasets with limited scale,and increasing the adaptability of deep learning models in scenarios with few training samples such as infrared object detection.Firstly,the ability of unlabeled samples in improving model generalization and suppressing overfitting under few labeled samples was described.Then,the process of semisupervised transfer learning for infrared object detection neural network was proposed:a pre-trained model was trained on large scale RGB dataset,and next it was fine-tuned using a few labeled and unlabeled IR images.Moreover,a pseudo-supervised loss function with feature similarity weighting was proposed,where the predictions from same batch was used as labels to each other,thus making full use of the feature distribution of similar objects in unlabeled images.To reduce the computation of semi supervised learning,the pseudosupervised loss of object was limited on the objects within the neighborhood of its feature vector.Experimental results show that the test accuracy of object detection neural network trained by proposed method is higher than that trained by supervised transfer learning,it achieves an improvement of 1.1%on Faster R-CNN and a significant improvement of 4.8%on YOLO-v3,which verifies the effectiveness of the proposed method.