بهبود استخراج راه در تصاویر پهپاد برمبنای یادگیری عمیق

Nowadays, road extraction is one of the most important analyses of remote sensing data. This process involves identifying and separating roads from other natural and artificial features in aerial or satellite images. According to growing demand for accurate and up-to-date information for urban planning, navigation system, traffic management, mapping, and monitoring, automatic and semi-automatic road extraction has become increasingly important. Although several research studies have been performed, there are still several challenges such as the proximity of trees, building shadows, similarity of road and sidewalk, and the presence of cars on roads.

In the recent decade, UAV platforms have been successfully applied in remote sensing applications, especially in remote sensing. According to the high spatial resolution of UAV images, they can be used in road extraction with high accuracy. Road extraction methods from 2D remote sensing images are divided into three categories: morphological operation, traditional machine learning algorithms, and deep learning-based methods. Although morphological methods are able to extract road shape features, they usually have little resistance to gaps, light changes, and contrast. Consequently, ML methods are proposed to solve these challenges. In this group, several textural features are manually extracted, and then classifiers are used for the final classification. The critical problem in road segmentation is accurately identifying pixels in an image as being on or off the road (background). The variety of road areas in terms of their location, size, form, and color makes developing effective segmentation algorithms more difficult. Additionally, when trees or buildings are covered by shadow in images, the accuracy of road segmentation is compromised. According to the high potential of the deep learning methods, they have been applied in the road extraction from remotely-sensed image.

We propose a framework to segment UAV images semantically, and the extracted roads are also refined by accurate spatial analysis around road boundaries. In this paper, a new post processing method is developed to enhance the result of convolutional neural networks (CNNs) by injection of spatial information. For this purpose, Xception as a powerful deep learning network is implemented to extract road in UAV images. Although, it detects road accurately but there are still false positive and negative pixels in the classification map. Consequently, post-processing step is proposed to decrease FP and FN by using spatial information. Several morphological operators are considered in this step and also road boundaries with their direction are computed. Fusion of the obtained results based on Xception and spatial information of the post-processing stage, significantly improve the road extraction results.

The proposed method has been evaluated on NITR and UDD datasets. The obtained results on NITR show that Xception reaches 93.81% where UNet and SegNet achieve 85.97% and 77.11%, respectively. On the UDD dataset, XCeption achieves 91.94% while UNet and SegNet reach 83.19% and 75.67%, respectively. The proposed method can improve the Xception results up to 96.08% and 94.47%.

The obtained results show that deep learning method uses spatial information in semantic segmentation, but the visual evaluation of the obtained classification map shows that there still exist misclassified pixels and more spatial analysis is required. Consequently, we proposed a post-processing step containing morphological operations and road boundaries analysis. The results on complex scenes show that the proposed post-processing step improves the recall metric up to 20%, but in less complex areas it improves less.

However, some limitations were also observed in the proposed method where the curved or irregular boundaries were not addressed. This indicates challenges in identifying more complex and irregular patterns, which may be due to structural limitations of the model or the inherent complexities of the data. Therefore, it is necessary to investigate and develop methods and optimize the algorithms in future research.