HDA-Net: Hybrid Convolutional Neural Networks for Small Objects Recognization at Airports

Abstract

Identifying small objects is essential in airport clearances, such as UAVs and bird flocks affecting the safety of the entire airport. Due to small objects and complex backgrounds, it is a formidable task to learn iconic representations from images. Convolution neural networks (CNNs) is the dominant framework in the past few years. Although CNNs can effectively mitigate local redundancy with convolutional kernels, it is also extremely challenging for convolutional kernels to obtain more feature information about the small object. Alternatively, the prediction box loss function has a critical influence on the eventual model performance. To resolve these problems, a revolutionary small object recognition network is proposed in this paper. In the structure, we propose a hybrid convolutional structure, which is able to obtain object features by different convolutional approaches to achieve more information about the small objects. Following that, we use the Double-One Path Aggregation Network (DO-PAN) feature fusion structure to further integrate the small object information. Finally, we propose all sides consistency intersection over union (AIoU) loss, which enables the model to predict the object location more exactly. Combining the three contributions, we designate this structure as HDA-Net. It can simply achieve 96.1% box mAP50 on the UAV-VIT dataset object recognition task, the inference speed on Tesla T4 is 28.46 fps.