Two new sequential search algorithms for feature selection in hyperspectral remote sensing images are proposed. Since
many wavebands in hyperspectral images are redundant and irrelevant, the use of feature selection to improve
classification results is highly needed. First, we present a new generalized steepest ascent (GSA) feature selection
technique that improves upon the prior steepest ascent algorithm by selecting a better starting search point and
performing a more thorough search. It is guaranteed to provide solutions that equal or exceed those of the classical
sequential forward floating selection algorithm. However, when the number of available wavebands is large, the
computational load required for the GSA algorithm becomes excessive. We thus propose a modification of the improved
floating forward selection algorithm which is more computationally efficient. Experimental results for two hyperspectral
data sets show that our proposed algorithms yield better classification results than other suboptimal search algorithms.
We address the important product inspection application of contaminant detection on chicken carcasses. Detection of four contaminant types of interest (duodenum, ceca, colon, and ingesta) from chickens fed with three different feeds (corn, milo, and wheat) is considered. We consider feature selection algorithms for choosing a small set of spectral bands (wavelengths) in hyperspectral (HS) data for online contaminant detection. For cases when an optimal solution is not realistic, we introduce our new improved forward floating selection algorithm; we call it a quasi-optimal (close to optimal) algorithm. Our algorithm is an improvement on the state-of-the-art sequential forward floating selection algorithm. We train our algorithm on a pixel database using only corn-fed chickens and test it on HS images of carcasses with three feeds. Our new algorithm gives an excellent detection rate and performs better than other suboptimal feature selection algorithms on this database.
We consider new methods to select useful sets of ratio features in hyperspectral data to detect contaminant regions on
chicken carcasses using data provided by ARS (Athens, GA). A ratio feature is the ratio of the response at each pixel
for two different wavebands. Ratio features perform a type of normalization and can thus help reduce false alarms, if a
good normalization algorithm is not available. Thus, they are of interest. We present a new algorithm for the general
problem of such feature selection in high-dimensional data. The four contaminant types of interest are three types of
feces from different gastrointestinal regions (duodenum, ceca, and colon) and ingesta (undigested food) from the
gizzard. To select the best two sets of ratio features from this 492-band HS data requires an exhaustive search of more
than seven billion combinations of two sets of ratio features, which is very excessive. Thus, we propose our new fast
ratio feature selection algorithm that requires evaluation of a much fewer number of sets of ratio features and is capable
of giving quasi-optimal or optimal sets of ratio features. This new feature selection method has not been previously
presented. It is shown to offer promise for an excellent detection rate and a low false alarm rate for this application. Our
tests use data with different feed types and different contaminant types.
Contaminant detection on chicken carcasses is an important product inspection application. The four contaminant types
of interest contain three types of feces from different gastrointestinal regions (duodenum, ceca, and colon) and ingesta
(undigested food) from the gizzard. Use of automated or semi-automated inspection systems for detecting fecal
contaminant regions is of great interest. Hyperspectral data provided by ARS (Athens, GA) were used to examine
detection of contaminants on carcasses. We address quasi-optimal algorithms for selecting a set of spectral bands
(wavelengths) in hyperspectral data for on-line contaminant detection (feature selection). We introduce our new
improved forward floating selection (IFFS) algorithm and compare its performance to that of other state-of-the-art
feature selection algorithms. Our initial results indicate that our method gives an excellent detection rate and performs
better than other feature selection algorithms. We also show that combination feature selection algorithms perform
worse.
Reduction of the potential health risks to consumers caused by food-borne infections is a very important food safety
issue of public concern; one of the leading causes of food-borne illnesses is fecal contamination. We consider detecting
fecal contaminants on chicken carcasses using hyperspectral imagery. We introduce our new improved floating forward
selection (IFFS) algorithm for feature selection of the wavebands to use in hyperspectral data for classification. Our
IFFS algorithm is an improvement on the state-of-the-art sequential floating forward selection (SFFS) algorithm. Our
initial results indicate that our method gives an excellent detection rate and performs better than other quasi-optimal
feature selection algorithms.
Detection of skin tumors on chicken carcasses is considered. A chicken skin tumor consists of an ulcerous lesion region
surrounded by a region of thickened-skin. We use a new adaptive branch-and-bound (ABB) feature selection algorithm
to choose only a few useful wavebands from hyperspectral data for use in a real-time multispectral camera. The ABB
algorithm selects an optimal feature subset and is shown to be much faster than any other versions of the branch and
bound algorithm. We found that the spectral responses of the lesion and the thickened-skin regions of tumors are
considerably different; thus we train our feature selection algorithm to separately detect the lesion regions and
thickened-skin regions of tumors. We then fuse the two HS detection results of lesion and thickened-skin regions to
reduce false alarms. Initial results on six hyperspectral cubes show that our method gives an excellent tumor detection
rate and a low false alarm rate.
We propose a new adaptive branch and bound (ABB) algorithm for selecting the optimal subset of features in hyperspectral applications. The algorithm improves the search speed by avoiding unnecessary criterion function calculations at nodes in the solution tree. Our algorithm includes the following new properties: (i) ordering the tree nodes by the significance of features during construction of the tree, (ii) obtaining a large "good" initial bound by a floating search method, (iii) a new method to select an initial starting search level in the tree, and (iv) a new adaptive jump search strategy to select subsequent search levels to avoid redundant criterion function calculations. Our experimental results for two databases demonstrate that our method is significantly faster than other versions of the branch and bound algorithm.
Detection is one of the most formidable problems in automatic target recognition, since it involves locating multiple classes of targets of interest with distortions present in cluttered scenes. Fast and efficient algorithms are needed for detection, since in detection we need to analyze every local region of large image scenes. Minimum noise and correlation energy (MINACE) filters are attractive distortion-invariant filters (DIFs); we consider MINACE filter use in detection, since they provide sharp correlation peak values for targets and overcome the effect of aspect view distortions in the input data. Most prior work on DIFs considered classification, not detection. MINACE filters seem to require fewer filters than do other DIFs, and they recognize objects with aspect views different by 15° from those present in the training set. They are also shift-invariant and require only a few filters to handle detection of multiple target classes. We test our improved MINACE filters to detect 8 classes of objects in an infrared (IR) database with a ±90° range of aspect views. Initial test results are excellent with only 3 filters needed and very low false alarm rates.
We describe a fast method for dimensionality reduction and feature selection of ratio features for classification in hyperspectral data. The case study chosen is to discriminate internally damaged almond nuts from normal ones. For this case study, we find that using the ratios of the responses in several wavebands provides better features than a subset of waveband responses. We find that use of the Euclidean Minimum Distance metric gives slightly better results than the more conventional Spectral Angle Mapper distance metric in a nearest neighbor classifier.
We consider using minimum noise and correlation energy (Minace) filters to detect objects in high-resolution Electro-Optical (EO) visible imagery. EO data is a difficult detection problem because only primitive features such as edges and corners are useful. This occurs because the targets and the background in EO data can have very similar gray levels, which leads to very low contrast targets; no hot spots (present in infrared (IR) data) or bright reflectors (present in synthetic aperture radar (SAR) data) exist in EO data. Since only geometrical (aspect view) distortions are expected in EO data (no thermal variations, as in IR, are expected), we consider using distortion-invariant Minace filters to detect targets. Such filters are shift-invariant and have been shown to be suitable for detection in other data (IR and SAR). Minace filters are attractive distortion-invariant filters (DIFs) because they require only a few filters to handle detection of multiple target classes. These filters must be modified for use on EO data. For EO data, zero-mean Minace filters formed from zero-mean, unit-energy data are used, and thus use of local zero-mean normalized correlations are needed. They show excellent initial detection results.
We consider a feature selection method to detect skin tumors on chicken carcasses using hyperspectral reflectance
data. This allows for faster data collection than does fluorescence data. A chicken skin tumor is an ulcerous lesion
region surrounded by a region of thickened-skin. Detection of chicken tumors is a difficult detection problem because
the tumors vary in size and shape; some tumors appear on the side of the chicken. In addition, different areas of normal
chicken skin have a variety of hyperspectral response variations, some of which are very similar to the spectral
responses of tumors. Similarly, different tumors and different parts of a tumor have different spectral responses. Thus,
proper classifier training is needed and many false alarms are expected. Since the spectral responses of the lesion and
the thickened-skin regions of tumors are considerably different, we train our feature selection algorithm to detect lesion
regions and to detect thickened-skin regions separately; we then process the resultant images and we fuse the two HS
detection results to reduce false alarms. Our new forward selection and modified branch and bound algorithm is used to
select a small number of λ spectral features that are useful for discrimination. Initial results show that our method offers
promise for a good tumor detection rate and a low false alarm rate.
We consider new distortion-invariant filters (DIFs) to detect objects in high-resolution Electro-Optical (EO) visible imagery. EO data is a difficult detection problem, because only primitive features such as edges and corners are useful. No hot spots (present in IR data) or bright reflectors (present in SAR data) exist in EO data. We thus expect many false alarms when we try to detect objects in EO data. We use new eigen-detection filters because they are shift-invariant, require only few filters and can handle multiple target classes. Initial results show that our filters, when using zero-mean data, perform well on EO data.
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.