In this paper, we focus on tackling the problem of automatic accurate localization of detected objects in high-resolution remote sensing images. The two major problems for object localization in remote sensing images caused by the complex context information such images contain are achieving generalizability of the features used to describe objects and achieving accurate object locations. To address these challenges, we propose a new object localization framework, which can be divided into three processes: region proposal, classification, and accurate object localization process. First, a region proposal method is used to generate candidate regions with the aim of detecting all objects of interest within these images. Then, generic image features from a local image corresponding to each region proposal are extracted by a combination model of 2-D reduction convolutional neural networks (CNNs). Finally, to improve the location accuracy, we propose an unsupervised score-based bounding box regression (USB-BBR) algorithm, combined with a nonmaximum suppression algorithm to optimize the bounding boxes of regions that detected as objects. Experiments show that the dimension-reduction model performs better than the retrained and fine-tuned models and the detection precision of the combined CNN model is much higher than that of any single model. Also our proposed USB-BBR algorithm can more accurately locate objects within an image. Compared with traditional features extraction methods, such as elliptic Fourier transform-based histogram of oriented gradients and local binary pattern histogram Fourier, our proposed localization framework shows robustness when dealing with different complex backgrounds.

Orientated object detection in aerial images is still a challenging task due to the bird’s eye view and the various scales and arbitrary angles of objects in aerial images. Most current methods for orientated object detection are anchor-based, which require considerable pre-defined anchors and are time consuming. In this article, we propose a new one-stage anchor-free method to detect orientated objects in per-pixel prediction fashion with less computational complexity. Arbitrary orientated objects are detected by predicting the axis of the object, which is the line connecting the head and tail of the object, and the width of the object is vertical to the axis. By predicting objects at the pixel level of feature maps directly, the method avoids setting a number of hyperparameters related to anchor and is computationally efficient. Besides, a new aspect-ratio-aware orientation centerness method is proposed to better weigh positive pixel points, in order to guide the network to learn discriminative features from a complex background, which brings improvements for large aspect ratio object detection. The method is tested on two common aerial image datasets, achieving better performance compared with most one-stage orientated methods and many two-stage anchor-based methods with a simpler procedure and lower computational complexity.

Object detection in very-high-resolution (VHR) remote sensing imagery remains a challenge. Environmental factors, such as illumination intensity and weather, reduce image quality, resulting in poor feature representation and limited detection accuracy. To enrich the feature representation and mine the underlying context information among objects, this article proposes a context-aware convolutional neural network (CA-CNN) model for object detection that includes proposal generation, context feature extraction, feature fusion, and classification. During feature extraction, we propose integrating a context-regions-of-interests (Context-RoIs) mining layer into the CNN model and extracting context features by mapping Context-RoIs mined from the foreground proposals to multilevel feature maps. Finally, the context features extracted from multilevel layers are fused into a single layer, and the proposals represented by the fused features are classified by a softmax classifier. In this article, through numerous experiments, we thoroughly explore the influence of key factors, such as Context-RoIs, different feature scales, and different spatial context window sizes. Because of the end-to-end network design approach, our proposed model simultaneously maintains high efficiency and effectiveness. We conducted all model testing on the public NWPU VHR-10 data set. The experimental results demonstrate that our proposed CA-CNN model achieves significantly improved model performance and better detection results compared with the state-of-the-art methods.

Because of recent advances in Convolutional Neural Networks (CNNs), traditional CNNs have been employed to extract thousands of codes as feature representations for image retrieval. In this paper, we propose that more powerful features for high-resolution remote sensing image representations can be learned using only several tens of codes; this approach can improve the retrieval accuracy and decrease the time and storage requirements. To accomplish this goal, we first investigate the learning of a series of features with different dimensions using a few tens to thousands of codes via our improved CNN frameworks. Then, a Principal Component Analysis (PCA) is introduced to compress the high-dimensional remote sensing image feature codes learned by traditional CNNs. Comprehensive comparisons are conducted to evaluate the retrieval performance based on feature codes of different dimensions learned by the improved CNNs as well as the PCA compression. To further demonstrate the powerful ability of the low-dimensional feature representation learned by the improved CNN frameworks, a Feature Weighted Map (FWM), which can perform feature visualization and provides a better understanding of the nature of Deep Convolutional Neural Networks (DCNNs) frameworks, is explored. All the CNN models are trained from scratch using a large-scale and high-resolution remote sensing image archive, which will be published and made available to the public. The experimental results show that our method outperforms state-of-the-art CNN frameworks in terms of accuracy and storage.

Automatically detecting airports from remote sensing images has attracted significant attention due to its importance in both military and civilian fields. However, the diversity of illumination intensities and contextual information makes this task difficult. Moreover, auxiliary features both within and surrounding the regions of interest are usually ignored. To address these problems, we propose a novel method that uses a multiscale fusion feature to represent the complementary information of each region proposal, which is extracted by constructing a GoogleNet with a light feature module model that has an additional light fully connected layer. Then, the fusion feature is input to a support vector machine whose performance is enhanced using a hard negative mining method. Finally, a simplified localization method is applied to tackle the problem of box redundancy and to optimize the locations of airports. An experiment demonstrates that the fusion feature outperforms other features on airport detection tasks from remote sensing images containing complicated contextual information.

Road segmentation in remote sensing images has been widely used in many fields. Semantic segmentation, based on deep learning, has become a hot topic for road segmentation. With the deepening of convolutional neural network (CNN) structures, features in the convolution layer that has more semantic information become more important for road segmentation. However, the spatial resolution of the convolutional layer reduced as the CNN network deepens, which causes the extracted roads to lose some important location information. To solve this problem, this letter proposes a novel end-to-end road segmentation method to effectively utilize the different levels of convolutional layers to enhance the model's ability to precisely perceive road edges and shapes. The model includes an encoder and a decoder. The encoder encodes the image to obtain the features of different levels and scales. The decoder consists of two modules: scale fusion module and scale sensitive module. In the scale fusion module, features in pooling layers of different scales are fused to obtain a fusion feature. In a scale sensitive module, a weight tensor at the end of the network is learned to evaluate the importance of fusion features. This road segmentation network has been experimentally verified using public data sets, which greatly improves the road segmentation accuracy and achieves good performance.