This work presents a neural network model capable of recognizing small and tiny objects in thermal images collected by unmanned aerial vehicles. Our model consists of three parts, the backbone, the neck, and the prediction head. The backbone is developed based on the structure of YOLOv5 combined with the use of a transformer encoder at the end. The neck includes a BI-FPN block combined with the use of a sliding window and a transformer to increase the information fed into the prediction head. The prediction head carries out the detection by evaluating feature maps with the Sigmoid function. The use of transformers with attention and sliding windows increases recognition accuracy while keeping the model at a reasonable number of parameters and computation requirements for embedded systems. Experiments conducted on public dataset VEDAI and our collected datasets show that our model has a higher accuracy than state-of-the-art methods such as ResNet, Faster RCNN, ComNet, ViT, YOLOv5, SMPNet, and DPNetV3. Experiments on the embedded computer Jetson AGX show that our model achieves a real-time computation speed with a stability rate of over 90%.

翻译：本文提出一种能够识别无人机采集热成像图像中小型及微小目标的神经网络模型。该模型由三部分组成：主干网络、颈部模块和预测头。主干网络基于YOLOv5结构开发，并在其末端结合使用Transformer编码器。颈部模块包含BI-FPN结构，同时结合滑动窗口与Transformer以增强输入至预测头的信息。预测头通过Sigmoid函数评估特征图来完成检测。引入注意力机制与滑动窗口的Transformer在保持嵌入式系统合理参数量与计算需求的同时提升了识别精度。在公开数据集VEDAI及自建数据集上的实验表明，本模型较ResNet、Faster RCNN、ComNet、ViT、YOLOv5、SMPNet及DPNetV3等现有最优方法具有更高准确率。在嵌入式计算机Jetson AGX上的实验显示，本模型可实现实时计算速度且稳定率达90%以上。