Object Detection in Aerial Images Using Feature Fusion Deep Networks

Abstract

Object detection acts as an essential part in a wide range of measurement systems in traffic management, urban planning, defense, agriculture, and so on. Convolutional Neural Networks-based researches reach a great improvement on detection tasks in natural scene images enjoying from the strong ability of feature representations. However, because of the high density, the small size of objects, and the intricate background, the current methods achieve relatively low precision in aerial images. The intention of this work is to obtain better detection performance in aerial images by designing a novel deep neural network framework called Feature Fusion Deep Networks (FFDN). The novel architecture combines a designed structural learning layer based on a graphical model. As a result, the network not only provides powerful hierarchical representation but also strengthens the spatial relationship between the high-density objects. We demonstrate the great improvement of the proposed FFDN on the UAV123 data set and another novel challenging data set called UAVDT benchmark. The objects which appear with small size, partial occlusion and out of view, as well as in the dark background can be detected accurately. INDEX TERMS Convolutional neural networks (CNNs), aerial images, feature fusion deep networks (FFDN), object detection.

Improved Kiwifruit Detection Using Pre-Trained VGG16 With RGB and NIR Information Fusion

“…The Feature-Fusion mode has a quite close AP (90.5%) to that of the Image-Fusion mode, although double convolutional layers were used to extract kiwifruit image features. There have been some studies working on the relationship between extracted features and detection results [38], [39]. Based on those researches, a hypothesis that the Feature-Fusion mode has two VGG16 networks to learn the features from the RGB and NIR images respectively, which results in a duplicate features learning of some important features belonging to both RGB and NIR images, such as fruit shape and calyx shape.…”

Lightweight Attention Pyramid Network for Object Detection and Instance Segmentation

“…Using residual learning to optimize the convolutional layer, and the model to fuse the features of different layers, the method achieved better results than other classifiers. To solve the complex situation such as the small size of the object in the aerial image, a feature fusion deep network was proposed [27], and the spatial relationship between objects was increased by the fusion of network layer features, thus accurate detection results were obtained. Facial expressions are very important information in human behavior research.…”

Vehicle Detection in Aerial Images Based on 3D Depth Maps and Deep Neural Networks

“…However, detecting vehicles in aerial images both accurately and quickly is challenging. As aerial images are taken from altitude with a top-down view, vehicles appear relatively small, and a single image may contain many vehicles [10]. Moreover, other objects, shadows, and various patterns, such as road markings, can appear similar to vehicles [3].…”

“…Specifically, they perform less satisfactorily in the localization of small objects in a large scene [21]. In addition, training these networks generally demands a high computational cost, and the lack of wellannotated training data adds to the challenge [10], [22].…”