Allen Hanson | University of Massachusetts Amherst (original) (raw)

Papers by Allen Hanson

IEEE Transactions on Pattern Analysis and Machine Intelligence, 1984

segment some of the test images in our paper.' Both algorithms are histogram based and each attem... more segment some of the test images in our paper.' Both algorithms are histogram based and each attempts to improve the fidelity of the statistics by a different form of localization of the histogram statistics. However, there are cases where the recursive algorithm would not generate an acceptable segmentation. The basis for successful detection of hidden clusters is that they are detected in the recursive segmentation because

One important task in image interpretation is the process of understanding and identifying segmen... more One important task in image interpretation is the process of understanding and identifying segments of an image. Ascender II is a knowledge-based vision system in which the selection of IU algorithms and the fusion of information provided by them is combined in an efficient way. A major problem with many knowledge-based systems is that the knowledge base, control mechanism and knowledge sources are combined into a single intertwined system and the addition of new knowledge or change of domain requires a significant effort. In Ascender II, the knowledge base and control mechanism (reasoning subsystem) are independent of the knowledge sources (visual subsystem). This gives the system the flexibility to add or change knowledge sources with only minor changes in the reasoning subsystem. The reasoning subsystem is implemented using a set of hierarchical Bayesian networks which allows an incremental classification of a region. Experiments on three different data sets are presented using an initial implementation of the system focusing primarily on building reconstruction. shown in Figure 1. Ascender II has been designed as a general purpose vision system, although our initial effort has focused primarily on recognizing and reconstructing building from aerial images. Less effort has gone into the knowledge networks and IU processes necessary for other objects, such as open fields, parking lots and vehicles. Ascender II assumes that it has as input a set of focus-of-attention regions in the image data. These regions can be generated in a variety of ways, including human interaction and cues from other sources such as maps or other classified images. In the experiments described later, our primary goal was detecting and extracting buildings, so the initial regions were generated using the Ascender I system to detect 2D building footprints. In one of the experiments (Fort Benning), the regions were constructed using Ascender I and a classified SAR image. Once the regions are available, an intelligent control system based on Bayesian networks drives IU processes which extract and fuse 2D and 3D features into evidence for a specific set of class labels. Based on accumulated evidence, the system identifies regions as representing an instance of one of the generic object classes defined within the system and (when possible) constructs a coherent 3D model for each region (Collins et al. 1998). 2 THE REASONING SUBSYSTEM The goal of the reasoning subsystem is to accumulate evidence sufficient to determine a plausible identity of the region in terms of the object classes represented within the system. A-priori knowledge about objects and their relationships is captured in a hierarchical Bayesian network (see below). Bayesian networks have been successfully used in systems required to combine and propagate evidence for and against a particular hypothesis (Pearl 1988, Jensen 1996). In the Ascender II system, we use these networks to determine what evidence to collect and what to do with the evidence once it has been obtained. The Bayesian networks were developed using the HUGIN system (Andersen et al. 1989). Evidence is obtained by using the network to select appropriate IU processes which obtain relevant feature information from the image(s). A-priori knowledge, in the form of initial prior probabilities associated with each object class, is used to select an IU process to use in the initial step. Generally, the process initially selected is fairly generic and measures simple features. In the case of a building, the kind of features measured might include the evidence for a center roof line, the number of L and T junctions on the boundary of the region, etc. As evidence accumulates for a particular hypothesis (e.g. a building), the IU process can become much more complex (and presumably return better evidence). Once evidence has been obtained, it is combined Controller Belief Network Reasoning Subsystem Visual Subsystem

International Journal of Pattern Recognition and Artificial Intelligence, 1993

The role of visual registration in executing assembly tasks is examined. A framework to execute a... more The role of visual registration in executing assembly tasks is examined. A framework to execute assembly operations is presented in which assembly tasks are executed using a nite state machine that supervises the interaction between registration and control com-ponents. Temporal registration is achieved visually us-ing a camera pair by combining image based feature tracking and weak camera-object calibration. A peg-in-hole insertion operation is demonstrated. 1

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences

Salt marshes are intertidal ecosystems valuable for services including coastal protection and car... more Salt marshes are intertidal ecosystems valuable for services including coastal protection and carbon sequestration. Restoration of salt marshes is popular in this era of climate change and sea-level rise, especially in areas where marshes have been historically altered, including in the Bay of Fundy. Salt marsh restoration involves landcover change through time as a community of halophytic vegetation develops in the study area. Restoration sites are difficult to survey using traditional on-foot methods, and developing remote sensing methods to survey them would increase efficiency of monitoring. The purpose of our study was to assess the capability of UAV multispectral imagery to map landcovers in a salt marsh restoration site in the Musquash Estuary, New Brunswick, Canada. We used the Random Forests (RF) supervised classifier and validated our maps using field data. We also evaluated the importance of textural features by running two classifications, with and without textural features. The classification omitting textural features had lower classification and validation accuracies (96.29 % and 91.23 %, respectively) than the classification and validation accuracies obtained by including textural features (99.56 % and 96.84 %, respectively). Additional work is required to test our method in different locations and seasons.

Journal of Parallel and Distributed Computing, 1992

We present an interpretation system which utilizes world knowledge in the form of simple object h... more We present an interpretation system which utilizes world knowledge in the form of simple object hypothesis rules, and more complex interpretation strategies attached to object and scene schemata, to reduce the ambiguities in image measurements. These rules involve sets of partiahy redundant features each of which defines an area of feature space which represents a “vote” for an object. Convergent evidence from multiple interpretation strategies is organized by top-down control mechanisms in the context of a partial interpretation. One such strategy extends a kernel interpretation derived through the selection of object exemplars, which represent the most reliable image specific hypotheses of a general object class, resuhing in the extension of partial interpretations from islands of reliability.

Real-Time Parallel Computing, 1981

Over the past decade the range of image analysis applications has greatly broadened. Usually the ... more Over the past decade the range of image analysis applications has greatly broadened. Usually the first steps of processing involve the transformation of a large spatial array of pixels into a more compact description of the image in terms of visually distinct syntactic elements and their characteristics. These “low-level” processes may achieve the desired goal directly or may serve as the input to a further set of interpretation processes.

In this paper, we present an algorithm for extracting straight lines from a greylevel image. The ... more In this paper, we present an algorithm for extracting straight lines from a greylevel image. The algorithm is region-based, where each region is the underlying structure for one line, but only pixels in the region with highest gradient magnitude contribute to the line. The algorithm is optimized for both quality and speed, and time was optimized for runs on the low-level part of the heterogeneous Image Understanding Architecture (IUA), where it operates in nearly frame rate. Line output of the algorithm was input to a performance evaluation algorithm, whose results, together with timing experiments, show that the algorithm has a high performance/quality ratio as compared with two other algorithms. 1 This work was supported in part by Army Research Laboratory contract DAAL02-91-K-0047 1 Contents 1 Introduction 3 2 Related Work 5 2.1 The Plane Fit Algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : 5 2.2 The Principal Axis Algorithm : : : : : : : : : : : : : : : ...

Under the DARPA RADIUS program, the University of Massachusetts (UMass) developed techniques to a... more Under the DARPA RADIUS program, the University of Massachusetts (UMass) developed techniques to automatically populate a site model with 3-D building models extracted from multiple, overlapping images. The Automated Site Construction, Extension, Detection and Refinement (ASCENDER) system incorporates several key ideas. First, 3-D reconstruction is based on geometric features that remain stable under a wide range of viewing and lighting conditions. Second, rigorous photogrammetric camera models are used to describe the relationship between pixels in an image and 3-D locations in the scene, so that diverse sensor characteristics and viewpoints can be effectively exploited. Third, information is fused across multiple images for increased accuracy and reliability. Finally, known geometric constraints are applied whenever possible to increase the efficiency and reliability of the reconstruction process. Ascender was the primary deliverable of the UMass RADIUS effort and was delivered in ...

We present the issues involved in information obfuscation both for the embedding of information w... more We present the issues involved in information obfuscation both for the embedding of information within digital signals and the detection of the hidden information. More standard cryptographic techniques acheive security through an encoding so that an ...

Texture mapping has wide and important applications in visualization and virtual reality. Sur-fac... more Texture mapping has wide and important applications in visualization and virtual reality. Sur-face texture extraction from a single image su ers from perspective distortion, data de ciency, and corruption caused by shadows and occlusions. In this paper, a system is developed ...

This paper describes an algorithm that uses multi-scale Gaussian differential features (MGDFs) fo... more This paper describes an algorithm that uses multi-scale Gaussian differential features (MGDFs) for face recognition. Results on standard sets indicate at least 96% recognition accuracy, and a comparable or better performance with other well known techniques. The MGDF based technique is very general; its original application included similarity retrieval in textures, trademarks, binary shapes and heterogeneous gray-level collections.

We propose a classification technique for face expression recognition using AdaBoost that learns ... more We propose a classification technique for face expression recognition using AdaBoost that learns by selecting the relevant global and local appearance features with the most discriminating information. Selectivity reduces the dimensionality of the feature space that in turn results in significant speed up during online classification. We compare our method with another leading margin-based classifier, the Support Vector Machines (SVM) and identify the advantages of using AdaBoost over SVM in this context. We use histograms of Gabor and Gaussian derivative responses as the appearance features. We apply our approach to the face expression recognition problem where local appearances play an important role. Finally, we show that though SVM performs equally well, AdaBoost feature selection provides a final hypothesis model that can easily be visualized and interpreted, which is lacking in the high dimensional support vectors of the SVM.

More complex object-dependent interpretation strategies are represented in a procedural form with... more More complex object-dependent interpretation strategies are represented in a procedural form within the schema knowledge structures [8,25]. These local control strategies provide top-down control over the interpretation process. Partial interpretations are ,-e extended from "islands of reliability" as in the HEARSAY paradigm 17,161. General .0:0 5,. ."- .

RESTRICTIVE MARKINGS 2.. SECURITY CLASSIFICATION AUTHORITY 3. DIST RI BUTION/A VAI LASBILI TY OF ... more RESTRICTIVE MARKINGS 2.. SECURITY CLASSIFICATION AUTHORITY 3. DIST RI BUTION/A VAI LASBILI TY OF REPORT 2b.OECASSFICTIO/OONGPACI4GSHEOLEApproved for public release; Zb.ECLSSIICAIONDOWORAINGCHEULEdistribution unlimited. 4. PERFORMING ORGANIZATION REPORT NuMBER(S) 5. MONITORING ORGANIZATION REPORT NUMBER(S) kVOSR-T1R 3-17 6&. NAME OF PERFORMING ORGANIZATION b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION

Ia. REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS UNCLASSIFIED 2a. SECURITY CLASSIFICATI... more Ia. REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS UNCLASSIFIED 2a. SECURITY CLASSIFICATION AUTHORITY 3. DISTRIBUTION /AVAILABILITY OF REPORT 2b. DECLASSIFICATION/DOWNGRADING SCHEDULE Approved for public release; distribution is unlimited. 4. PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S) ETL-0549 6.. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION

IEEE Transactions on Pattern Analysis and Machine Intelligence, 1984

segment some of the test images in our paper.' Both algorithms are histogram based and each attem... more segment some of the test images in our paper.' Both algorithms are histogram based and each attempts to improve the fidelity of the statistics by a different form of localization of the histogram statistics. However, there are cases where the recursive algorithm would not generate an acceptable segmentation. The basis for successful detection of hidden clusters is that they are detected in the recursive segmentation because

One important task in image interpretation is the process of understanding and identifying segmen... more One important task in image interpretation is the process of understanding and identifying segments of an image. Ascender II is a knowledge-based vision system in which the selection of IU algorithms and the fusion of information provided by them is combined in an efficient way. A major problem with many knowledge-based systems is that the knowledge base, control mechanism and knowledge sources are combined into a single intertwined system and the addition of new knowledge or change of domain requires a significant effort. In Ascender II, the knowledge base and control mechanism (reasoning subsystem) are independent of the knowledge sources (visual subsystem). This gives the system the flexibility to add or change knowledge sources with only minor changes in the reasoning subsystem. The reasoning subsystem is implemented using a set of hierarchical Bayesian networks which allows an incremental classification of a region. Experiments on three different data sets are presented using an initial implementation of the system focusing primarily on building reconstruction. shown in Figure 1. Ascender II has been designed as a general purpose vision system, although our initial effort has focused primarily on recognizing and reconstructing building from aerial images. Less effort has gone into the knowledge networks and IU processes necessary for other objects, such as open fields, parking lots and vehicles. Ascender II assumes that it has as input a set of focus-of-attention regions in the image data. These regions can be generated in a variety of ways, including human interaction and cues from other sources such as maps or other classified images. In the experiments described later, our primary goal was detecting and extracting buildings, so the initial regions were generated using the Ascender I system to detect 2D building footprints. In one of the experiments (Fort Benning), the regions were constructed using Ascender I and a classified SAR image. Once the regions are available, an intelligent control system based on Bayesian networks drives IU processes which extract and fuse 2D and 3D features into evidence for a specific set of class labels. Based on accumulated evidence, the system identifies regions as representing an instance of one of the generic object classes defined within the system and (when possible) constructs a coherent 3D model for each region (Collins et al. 1998). 2 THE REASONING SUBSYSTEM The goal of the reasoning subsystem is to accumulate evidence sufficient to determine a plausible identity of the region in terms of the object classes represented within the system. A-priori knowledge about objects and their relationships is captured in a hierarchical Bayesian network (see below). Bayesian networks have been successfully used in systems required to combine and propagate evidence for and against a particular hypothesis (Pearl 1988, Jensen 1996). In the Ascender II system, we use these networks to determine what evidence to collect and what to do with the evidence once it has been obtained. The Bayesian networks were developed using the HUGIN system (Andersen et al. 1989). Evidence is obtained by using the network to select appropriate IU processes which obtain relevant feature information from the image(s). A-priori knowledge, in the form of initial prior probabilities associated with each object class, is used to select an IU process to use in the initial step. Generally, the process initially selected is fairly generic and measures simple features. In the case of a building, the kind of features measured might include the evidence for a center roof line, the number of L and T junctions on the boundary of the region, etc. As evidence accumulates for a particular hypothesis (e.g. a building), the IU process can become much more complex (and presumably return better evidence). Once evidence has been obtained, it is combined Controller Belief Network Reasoning Subsystem Visual Subsystem

International Journal of Pattern Recognition and Artificial Intelligence, 1993

The role of visual registration in executing assembly tasks is examined. A framework to execute a... more The role of visual registration in executing assembly tasks is examined. A framework to execute assembly operations is presented in which assembly tasks are executed using a nite state machine that supervises the interaction between registration and control com-ponents. Temporal registration is achieved visually us-ing a camera pair by combining image based feature tracking and weak camera-object calibration. A peg-in-hole insertion operation is demonstrated. 1

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences

Salt marshes are intertidal ecosystems valuable for services including coastal protection and car... more Salt marshes are intertidal ecosystems valuable for services including coastal protection and carbon sequestration. Restoration of salt marshes is popular in this era of climate change and sea-level rise, especially in areas where marshes have been historically altered, including in the Bay of Fundy. Salt marsh restoration involves landcover change through time as a community of halophytic vegetation develops in the study area. Restoration sites are difficult to survey using traditional on-foot methods, and developing remote sensing methods to survey them would increase efficiency of monitoring. The purpose of our study was to assess the capability of UAV multispectral imagery to map landcovers in a salt marsh restoration site in the Musquash Estuary, New Brunswick, Canada. We used the Random Forests (RF) supervised classifier and validated our maps using field data. We also evaluated the importance of textural features by running two classifications, with and without textural features. The classification omitting textural features had lower classification and validation accuracies (96.29 % and 91.23 %, respectively) than the classification and validation accuracies obtained by including textural features (99.56 % and 96.84 %, respectively). Additional work is required to test our method in different locations and seasons.

Journal of Parallel and Distributed Computing, 1992

We present an interpretation system which utilizes world knowledge in the form of simple object h... more We present an interpretation system which utilizes world knowledge in the form of simple object hypothesis rules, and more complex interpretation strategies attached to object and scene schemata, to reduce the ambiguities in image measurements. These rules involve sets of partiahy redundant features each of which defines an area of feature space which represents a “vote” for an object. Convergent evidence from multiple interpretation strategies is organized by top-down control mechanisms in the context of a partial interpretation. One such strategy extends a kernel interpretation derived through the selection of object exemplars, which represent the most reliable image specific hypotheses of a general object class, resuhing in the extension of partial interpretations from islands of reliability.

Real-Time Parallel Computing, 1981

Over the past decade the range of image analysis applications has greatly broadened. Usually the ... more Over the past decade the range of image analysis applications has greatly broadened. Usually the first steps of processing involve the transformation of a large spatial array of pixels into a more compact description of the image in terms of visually distinct syntactic elements and their characteristics. These “low-level” processes may achieve the desired goal directly or may serve as the input to a further set of interpretation processes.

In this paper, we present an algorithm for extracting straight lines from a greylevel image. The ... more In this paper, we present an algorithm for extracting straight lines from a greylevel image. The algorithm is region-based, where each region is the underlying structure for one line, but only pixels in the region with highest gradient magnitude contribute to the line. The algorithm is optimized for both quality and speed, and time was optimized for runs on the low-level part of the heterogeneous Image Understanding Architecture (IUA), where it operates in nearly frame rate. Line output of the algorithm was input to a performance evaluation algorithm, whose results, together with timing experiments, show that the algorithm has a high performance/quality ratio as compared with two other algorithms. 1 This work was supported in part by Army Research Laboratory contract DAAL02-91-K-0047 1 Contents 1 Introduction 3 2 Related Work 5 2.1 The Plane Fit Algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : 5 2.2 The Principal Axis Algorithm : : : : : : : : : : : : : : : ...

Under the DARPA RADIUS program, the University of Massachusetts (UMass) developed techniques to a... more Under the DARPA RADIUS program, the University of Massachusetts (UMass) developed techniques to automatically populate a site model with 3-D building models extracted from multiple, overlapping images. The Automated Site Construction, Extension, Detection and Refinement (ASCENDER) system incorporates several key ideas. First, 3-D reconstruction is based on geometric features that remain stable under a wide range of viewing and lighting conditions. Second, rigorous photogrammetric camera models are used to describe the relationship between pixels in an image and 3-D locations in the scene, so that diverse sensor characteristics and viewpoints can be effectively exploited. Third, information is fused across multiple images for increased accuracy and reliability. Finally, known geometric constraints are applied whenever possible to increase the efficiency and reliability of the reconstruction process. Ascender was the primary deliverable of the UMass RADIUS effort and was delivered in ...

We present the issues involved in information obfuscation both for the embedding of information w... more We present the issues involved in information obfuscation both for the embedding of information within digital signals and the detection of the hidden information. More standard cryptographic techniques acheive security through an encoding so that an ...

Texture mapping has wide and important applications in visualization and virtual reality. Sur-fac... more Texture mapping has wide and important applications in visualization and virtual reality. Sur-face texture extraction from a single image su ers from perspective distortion, data de ciency, and corruption caused by shadows and occlusions. In this paper, a system is developed ...

This paper describes an algorithm that uses multi-scale Gaussian differential features (MGDFs) fo... more This paper describes an algorithm that uses multi-scale Gaussian differential features (MGDFs) for face recognition. Results on standard sets indicate at least 96% recognition accuracy, and a comparable or better performance with other well known techniques. The MGDF based technique is very general; its original application included similarity retrieval in textures, trademarks, binary shapes and heterogeneous gray-level collections.

We propose a classification technique for face expression recognition using AdaBoost that learns ... more We propose a classification technique for face expression recognition using AdaBoost that learns by selecting the relevant global and local appearance features with the most discriminating information. Selectivity reduces the dimensionality of the feature space that in turn results in significant speed up during online classification. We compare our method with another leading margin-based classifier, the Support Vector Machines (SVM) and identify the advantages of using AdaBoost over SVM in this context. We use histograms of Gabor and Gaussian derivative responses as the appearance features. We apply our approach to the face expression recognition problem where local appearances play an important role. Finally, we show that though SVM performs equally well, AdaBoost feature selection provides a final hypothesis model that can easily be visualized and interpreted, which is lacking in the high dimensional support vectors of the SVM.

More complex object-dependent interpretation strategies are represented in a procedural form with... more More complex object-dependent interpretation strategies are represented in a procedural form within the schema knowledge structures [8,25]. These local control strategies provide top-down control over the interpretation process. Partial interpretations are ,-e extended from "islands of reliability" as in the HEARSAY paradigm 17,161. General .0:0 5,. ."- .

RESTRICTIVE MARKINGS 2.. SECURITY CLASSIFICATION AUTHORITY 3. DIST RI BUTION/A VAI LASBILI TY OF ... more RESTRICTIVE MARKINGS 2.. SECURITY CLASSIFICATION AUTHORITY 3. DIST RI BUTION/A VAI LASBILI TY OF REPORT 2b.OECASSFICTIO/OONGPACI4GSHEOLEApproved for public release; Zb.ECLSSIICAIONDOWORAINGCHEULEdistribution unlimited. 4. PERFORMING ORGANIZATION REPORT NuMBER(S) 5. MONITORING ORGANIZATION REPORT NUMBER(S) kVOSR-T1R 3-17 6&. NAME OF PERFORMING ORGANIZATION b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION

Ia. REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS UNCLASSIFIED 2a. SECURITY CLASSIFICATI... more Ia. REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS UNCLASSIFIED 2a. SECURITY CLASSIFICATION AUTHORITY 3. DISTRIBUTION /AVAILABILITY OF REPORT 2b. DECLASSIFICATION/DOWNGRADING SCHEDULE Approved for public release; distribution is unlimited. 4. PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S) ETL-0549 6.. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION