In this work, we report a performance comparison of an acousto-optic tunable filter (AOTF), and a liquid crystal tunable filter (LCTF) based on a novel dual-arm hyperspectral imaging (HSI) configuration. The main purpose of this work is to highlight the leverage points of each tunable filter, in order to facilitate filter choice in HSI design. Three main parameters are experimentally examined: spectral resolution, out-of-band suppression, and image quality in the sense of spatial resolution. The experimental results, using wideband illumination, laser lines, and a spatial test target (USAF-1951) emphasized the superiority of AOTF in spectral resolution, out-of –band suppression and random switching speed between wavelengths. The same experiments demonstrated LCTF to have better performance in terms of the spatial image resolution, both horizontal and vertical, and high definition quality. In conclusion, the efficient design of an HSI system is application-dependent. For medical applications, for instance, if the tissue of interest has undefined optical properties, or contains close spectral features, AOTF might be the better option. Otherwise, LCTF is more convenient and simpler to use, especially if the tissue chromophore’s spatial mapping is needed.
Early detection and treatment of high-grade dysplasia (HGD) in Barrett’s esophagus may reduce the risk of developing esophageal adenocarcinoma. Confocal endomicroscopy (CLE) has shown advantages over routine white-light endoscopic surveillance with biopsy for histological examination; however, CLE is compromised by insufficient contrast and by intra- and interobserver variation. An FDA-approved PDT photosensitizer was used here to reveal morphological and textural features similar to those found in histological analysis. Support vector machines were trained using the aforementioned features to obtain an automatic and robust detection of HGD. Our results showed 95% sensitivity and 87% specificity using the optimal feature combination and demonstrated the potential for extension to a three-dimensional cell model.
Advents in new sensing hardwares like GigE-cameras and fast growing data transmission capability create an imbalance
between the amount of large scale aerial imagery and the means at disposal for treating them. Selection of saliency
regions can reduce significantly the prospecting time and computation cost for the detection of objects in large scale
aerial imagery. We propose a new approach using multiscale Simple Linear Iterative Clustering (SLIC) technique to
compute the saliency regions. The SLIC is fast to create compact and uniform superpixels, based on the distances in both
color and geometric spaces. When a salient structure of the object is over-segmented by the SLIC, a number of
superpixels will follow the edges in the structure and therefore acquires irregular shapes. Thus, the superpixels
deformation betrays presence of salient structures. We quantify the non-compactness of the superpixels as a salience
measure, which is computed using the distance transform and the shape factor. To treat objects or object details of
various sizes in an image, or the multiscale images, we compute the SLIC segmentations and the salient measures at
multiple scales with a set of predetermined sizes of the superpixels. The final saliency map is a sum of the salience
measures obtained at multiple scales. The proposed approach is fast, requires no input of user-defined parameter,
produces well defined salient regions at full resolution and adapted to multi-scale image processing.
For detecting vehicles in large scale aerial images we first used a non-parametric method proposed recently by Rosin to
define the regions of interest, where the vehicles appear with dense edges. The saliency map is a sum of distance
transforms (DT) of a set of edges maps, which are obtained by a threshold decomposition of the gradient image with a
set of thresholds. A binary mask for highlighting the regions of interest is then obtained by a moment-preserving
thresholding of the normalized saliency map. Secondly, the regions of interest were over-segmented by the SLIC
superpixels proposed recently by Achanta et al. to cluster pixels into the color constancy sub-regions. In the aerial
images of 11.2 cm/pixel resolution, the vehicles in general do not exceed 20 x 40 pixels. We introduced a size constraint
to guarantee no superpixels exceed the size of a vehicle. The superpixels were then classified to vehicle or non-vehicle
by the Support Vector Machine (SVM), in which the Scale Invariant Feature Transform (SIFT) features and the Linear
Binary Pattern (LBP) texture features were used. Both features were extracted at two scales with two size patches. The
small patches capture local structures and the larger patches include the neighborhood information. Preliminary results
show a significant gain in the detection. The vehicles were detected with a dense concentration of the vehicle-class
superpixels. Even dark color cars were successfully detected. A validation process will follow to reduce the presence of
isolated false alarms in the background.
In the aerial images of 11.2 cm/pixel resolution the car components that can be seen are only large parts of the car such
as car bodies, windshields, doors and shadows. Furthermore, these components are distorted by low spatial resolution,
low color contrast, specular reflection and viewpoint variation. We use the mean shift procedure for robust segmentation
of the car parts in the geometric and color joint space. This approach is robust, efficient, repeatable and independent of
the threshold parameters. We introduce a hierarchical segmentation algorithm with three consecutive mean-shift
procedures. Each is designed with a specific bandwidth to segment a specific car part, whose size is estimated a priori,
and is followed by a support vector machine in order to detect this car part, based on the color features and the
geometrical moment based features. The procedure starts with the largest car parts, which are then removed from the
segmented region lists after the detection to avoid over-segmentation of large regions with the mean-shift using smaller
bandwidth values. Finally we detect and count the cars in the image by combining the detected car parts according to the
spatial relations. Experiment results show a good performance.
We propose a feature-based approach for vehicle detection in aerial imagery with 11.2 cm/pixel resolution.
The approach is free of all constraints related to the vehicles appearance. The scale-invariant feature
transform (SIFT) is used to extract keypoints in the image. The local structure in the neighbouring of the
SIFT keypoints is described by 128 gradient orientation based features. A Support Vector Machine is used
to create a model which is able to predict if the SIFT keypoints belong to or not to car structures in the
image. The collection of SIFT keypoints with car label are clustered in the geometric space into subsets and
each subset is associated to one car. This clustering is based on the Affinity Propagation algorithm
modified to take into account specific spatial constraint related to geometry of cars at the given resolution.
The context-driven target recognition requires the object-of-interest (OOI) to be first detected. We use the multiscale beamlet transform to detect airport runways as the OOI for detecting the aircraft. The up-to-down strategy in the beamlet graph structure is used for the connectivity and directional continuation of the edges, which are first detected in a coarse scale and are then refined at several finer scales.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.