In this chapter a novel feature map was introduced, which is able to improve the efficiency of parametric active contour methods (GVF and VFC) when detecting complex boundaries with weak contrast and high curvatures. This feature map is based on a modified Harris characteristic function and describes better the principal curvatures and emphasizes both corners and edges of the object equally.
By calculating local maxima of the feature map, a point set is given, which can be also applied successfully in other applications for complex object detection,
like searching for MS lesions in brain MRI image pairs or localization of small flying targets.
Although the method has been extended to multiple object detection, detailed evaluation is required to analyze the efficiency with more complex images, like in case of background clutter. It would be also interesting to extend the method to 3-D. However, the three dimensional representation of Harris corner detector is quite novel [85] and needs more extensive research and evaluation.
Chapter 4
Improved Harris Feature Point Set for Orientation Sensitive Detection in Aerial Images
This chapter addresses automatic detection tasks (urban area and building de-tection) in remotely sensed images. As manual administration is time consuming and unfeasible, researchers have to focus on automated processing techniques, which can handle various image characteristics and huge amount of data. The introduced method extracts feature points in the first step, which is followed by different detection techniques concentrating on urban areas or buildings. Orien-tation information in the close proximity of the feature points is exploited for constructing a more accurate voting map to represent urban areas [3] and to ex-tract an orientation sensitive edge/connectivity map, emphasizing edges only in the main directions to detect buildings [11]. This chapter presents methodologi-cal contributions in two key issues of the detection process: (1) An automatimethodologi-cally extracted, Harris based feature point set is introduced for the first step, which is able to represent urban areas more precisely. (2) Orientation information in the local neighborhood of points is exploited and applied as a new feature for more accurate urban area and building detection. Evaluation results show that the proposed contributions increase the detection accuracy of both urban areas and buildings.
81
4.1 Introduction
Automatic detection of urban areas and/or buildings in optical aerial images means a great support in urban development analysis, map updating, disaster management and also helps municipalities in long-term residential area planning.
Unfortunately, the images may cover large areas and they can be taken in altering weather and illumination conditions, causing very different image features. More-over, as urban areas are usually changing dynamically, continuous administration is required to have up-to-date information. Since manual administration is time consuming and unfeasible, researchers have to focus on automated processing techniques, which can handle various image characteristics and huge amount of data.
A wide range of automatic urban (a.k.a. built-in) area detection techniques have been introduced in recent years (overview of literature about automatic building detection can be read in Section 3.5.1.1). The first group of such meth-ods needs specific training data to detect urban areas, like [91] using a differential morphological profile to record structural image information, then applied feature extraction and neural network for classifying the features; and [39] presenting a multiple conditional random fields ensemble model to incorporate multiple fea-tures and learn their contextual information.
This work follows another methodology, to construct a direct method which does not need any training data for urban area extraction. [92] showed that cor-ner detectors (Harris [24] and SUSAN [100]) are efficient tools for distinguishing different types of structures (man-made versus natural structures) present in the image. Since then, other works also applied interest point detectors for urban area detection: [67] used SIFT (Section 2.2.1) integrated with graph theory. The results were promising, but the computational complexity and time was quite significant. To reduce computational requirements, the same authors introduced a novel technique using Gabor feature points and spatial voting [88].
In this chapter, first an automatically extracted feature point set, called Mod-ified Harris for Edges and Corners (MHEC) is introduced for urban area detec-tion, which was introduced for effective object contour detection in Section 3.3.2 and used in different applications (see Section 3.5). As the density of extracted
Figure 4.1: Simplified diagram of the workflow of urban area detection.
feature points is higher in the residential areas [92], building a probability map based on this feature point set can help in identifying urban areas and buildings.
After having a local feature point set, the voting matrix strategy of [88] is applied to get a probability map of urban area. To improve the accuracy of this step, a novel orientation-sensitive technique is also proposed for constructing the voting matrix. Finally, urban areas are calculated by a decision-making step. Figure 4.1 shows the main algorithmic steps of the method. In the experimental part, it is demonstrated that the introduced method is able to outperform other inter-est point detectors, moreover the proposed orientation-sensitive voting matrix is also able to improve the performance of the previously introduced Gabor-based algorithm [88] as well.
Orientation sensitivity also exists in case of building detection. According to our assumption, a small urban area has buildings with connected orientation.
In most cases, houses are oriented according to some bigger structure (e. g. the road network), therefore main orientation of the area can be defined. As the proposed MHEC point set represents the area efficiently, the idea was to calculate the main direction of the buildings of the urban area based on this point set.
Orientation information of the points is calculated based on the edges of the local neighborhood, and main directions in an urban region are determined with statistical Gaussian function fitting.
After defining the main orientation(s), edges in the given direction have to be enhanced and for this issue, shearlet approach [93] is used. Shearlet is a multidimensional version of the traditional wavelet transform. Shearlets, unlike wavelets, are theoretically optimal in representing images with edges and have the ability to fully capture directional features. This technique was used to generate a connectivity map where main directional edges are emphasized.
Based on this connectivity map, the feature point set is divided into subsets, representing building candidates. Finally, Chan-Vese active contour method [47]
estimates boundary of the building.