Recently image processing such as noise reduction, restoration, and super-resolution using the soft-partition-based weighted sum filters have shown state-of-the-art results. The partition-based weighted sum filters are spatially adaptive filtering techniques by combining vector quantization and linear finite impulse response filtering, which have been shown to achieve much better results than spatial-invariant filtering methods. However, they are computationally prohibitive for practical applications because of enormous computation involved in both filtering and training. Real-time filtering is impossible even for small image and window sizes. This paper presents fast implementations of the soft-partition-based weighted sum filtering by exploiting the massively parallel processing capabilities of a GPU within the CUDA framework. For the implementations, we focus on memory management and implementation strategies. The performance on various image and window sizes is measured and compared between the GPU-based and CPU-based implementations. The results show that the GPU-based implementations can significantly accelerate computations for the soft-partition-based weighted sum filtering, and make real-time image filtering possible.
In this paper we present an object boundary detection system using an off-the-shelf available 3D stereo monitor. Instead of implementing algorithms, the system’s image processing is based on utilizing the polarization feature of liquid-crystal display and the way the image is displayed on the 3D monitor to enhance object boundary. The users can view the enhanced object contour through a polarization glasses in real-time, which can be also recorded using a camera for further processing. A software is developed for user interaction and providing feedback to obtain the best detection results. The effectiveness of the proposed system is demonstrated using some medical and biological images. The proposed system has the advantages of real-time high speed processing, almost no numerical computation, and robustness to noise over the traditional methods using image processing algorithms.
Active contours, as a technique for boundary extraction, have been successfully used in image processing and computer vision. One of the knotty problems of active contours is to conform to the object boundary with complex shape, which could bring heavy manual work at the initialization procedure. The gradient vector flow (GVF) field has been one of the most popular external forces that can increase the capture range of active contours and bidirectionally evolve the active contours toward the object boundary. However, it has a poor performance when dealing with some complex shapes, such as semi-closed concave, screwy concave, hooked concave, as well as the others presented in our experiments. We propose a novel GVF-based balloon force, which can efficiently assist the GVF field in driving active contours toward the complex object shapes. This additional force is used only when the active contours are prevented from evolving toward the object boundary by the saddle and/or stationary points in the GVF field. Therefore, it can maintain the bidirectional evolution property of the GVF and meanwhile take advantage of the power of the balloon force in segmenting complex shapes. Various experimental results on image segmentation are presented to show the good performance of the proposed active contour model that uses the GVF field and the proposed balloon force together.
We present a new compressed domain method for tracking objects in airborne videos. In the
proposed scheme, a statistical snake is used for object segmentation in I-frames, and motion
vectors extracted from P-frames are used for tracking the object detected in I-frames. It is shown
that the energy function of the statistical snake can be obtained directly from the compressed
DCT coefficients without the need of full decompression. The number of snake deformation
iterations can be also significantly reduced in compressed domain implementation. The
computational cost is significantly reduced by using compressed domain processing while the
performance is competitive to that of pixel domain processing. The proposed method is tested
using several UAV video sequences, and experiments show that the tracking results are
satisfactory.
By incorporating the image gradient directional information into the geodesic active contour model, we propose a novel active contour model called directional geodesic active contour, which has the advantage of selectively detecting the image edges with different gradient directions. The experiment results show the high performance of the proposed active contour in image segmentation, especially when multiple edges with different gradient directions are present near the object boundary to confuse the active contour.
Fusing acoustic and visual data has the potential to significantly improve target tracking performance by exploiting
the complementary and redundant information. However, it is not clear how to efficiently fuse these two modalities
and how to reduce the high energy consumption of video sensors to improve network lifespan. High energy
requirement of video sensors is mainly due to the high sample acquisition cost and high computational complexity
of target detection algorithms. We address the computation energy by performing target detection in a region-ofinterest
and studying target detection algorithms' computational complexity and data fusion architecture. Some
strategies are proposed to help video sensors in performing target detection and reducing the power consumption by
incorporating the acoustic detection result. We study two widely used target detection algorithms, background
subtraction and snake, for their use in joint acoustic-video tracking, and analyze their computational complexities
and memory requirements. We hope that the analysis can effectively guide the designer to select the most suitable
algorithm and architecture for a given scenario or application.
A novel binary level set method for boundary-based image segmentation is proposed, which is extended from region-based binary level set methods. The proposed binary level set method is based on the geometric active contour framework, which is a traditional level set method applied in boundary-based image segmentation. However, being different from the geometric active contour, the proposed binary level set method replaces the traditional level set function with a binary level set function to reduce the expensive computational cost of redistancing the traditional level set function. The experiments and complexity analysis show that the proposed binary level set method is more efficient than the geometric active contour for image segmentation while giving similar results to the geometric active contour.
We present a modified binary level set method for two-phase image segmentation, which is based on the binary level set method originally proposed by X.-C. Tai et al. [Int. J. Comput. Vis. 73, 61–76 (2007)]. The modified binary level set method is superior to Tai et al.'s binary level set method in preserving the curve evolution property of gradually shrinking and expanding the partition interface represented by a binary level set function in the segmentation process. Some experiments conducted on real images show the good results of the modified binary level set method.
The high computational complexity of level set methods has excluded themselves from many real-time applications. The high algorithm complexity is mainly due to the need of solving partial differential equations (PDEs) numerically. For image segmentation and object tracking applications, it is possible to approximate level set curve evolution process without solving PDEs since we are interested in the final object boundary instead of the accurate curve evolution process. This paper proposes a fast parallel method to simplify curve evolution process using simple binary morphological operations. The proposed fast implementation allows real-time image segmentation and object tracking using level set curve evolution, while preserves the advantage of level set methods for automatically handling topological changes. It can utilize the parallel processing capability of existing embedded hardware, parallel computers or optical processors for fast curve evolution.
Super-resolution reconstruction algorithms have been demonstrated to be very effective in enhancing image spatial resolution by combining several low-resolution images to yield a single high-resolution image. However, the high computational complexity has become a major obstacle for the use of super-resolution techniques in real time applications. Most previous computationally efficient super-resolution techniques have been focused on reducing the number of iterations due to the iterative nature of most super-resolution algorithms. In this paper, we propose a region-of-interest (ROI) image preprocessing technique to improve the processing speed of super-resolution reconstruction. To better integrate the preprocessing with super-resolution, the proposed ROI extraction technique is developed under the same statistical framework as super-resolution. Simulation results are provided to demonstrate the performance of the proposed method.
Target detection and tracking in real-time videos are very important and yet difficult for many applications. Numerous detection and tracking techniques have been proposed, typically by imposing some constraints on the motion and image to simplify the problem depending on the application and environment. This paper focuses on target detection and tracking in airborne videos, in which not much simplification can be made. We have recently proposed a combined/switching detection and tracking method which is based on the combination of a spatio-temporal segmentation and statistical snake model. This paper improves the statistical snake model by incorporating both edge and region information and enhancing the snake contour deformation. A more complex motion model is used to improve the accuracy of object detection and size classification. Mean-shift is integrated into the proposed combined method to track small point objects and deal with the problem of object disappearance-reappearance. Testing results using real UAV videos are provided.
Digital hologram compression has recently received increasing attention due to easy acquisition and new applications in three-dimensional information processing. Standard compression algorithms perform poorly on complex-valued holographic data. This paper studies quantization techniques for lossy compression of digital holographic images, where three commonly used quantizers are compared. Our observations show that the real and imagery components of holograms and their corresponding Fourier transform coefficients exhibit a Laplacian and Gaussian distribution, respectively. It is therefore possible to design an optimal quantizer for holographic data compression. To further increase the compression ratio, preprocessing techniques to extract the region of interest are presented. These include Fourier plane filtering and statistical snake image segmentation.
Level set methods have recently been applied to a number of image processing and computer vision applications. However, they have been suffered from the computational complexity. This paper exploits the parallel processing capability of opto-electronic signal processing systems for real-time level-set computation. The most timeconsuming
operation in level set methods is the computation of the signed distance function from each grid point to the zero-level set. We show that it can be represented by thresholded correlation operations, and therefore it can be easily realized using an optical correlator. A fast level-set image processing system is constructed based on a joint
transform correlation architecture. The simulation results are given.
This paper describes a system for tracking objects in a video stream obtained from a moving airborne platform, which is applied to annotate video objects automatically. The object to annotate is indicated by a mouse click. The proposed tracking algorithm uses a spatio-temporal segmentation followed by temporal tracking. It differs from existing techniques by the following features. The same algorithm is used in tracking both moving and stationary objects by making the stationary objects "move." It is general enough to handle any objects of various types and sizes including point objects. The system has a fast implementation because all image operations are applied on small image regions only. The effectiveness of the proposed algorithm is demonstrated on a few real video sequences.
Phase shifting error is a major error source in phase-shifting digital holography. It affects the quality of the reconstructed object and causes errors in its phase and amplitude calculation. This paper presents a simple method to accurately retrieve the actual phase-shifting value used in practical hologram recording. It therefore provides the possibility of completely eliminating the phase-shifting error in digital hologram. The proposed method is based on solving the object wave reconstruction equation. The effectiveness of the proposed method is verified by both mathematical analysis and computer simulation.
This paper describes a system for tracking objects in a video stream obtained from a moving airborne platform. The system is applied to annotate video objects automatically. The object to annotate is indicated by a mouse click. The proposed tracking algorithm is based on the general framework of spatiotemporal segmentation followed by temporal tracking. However, it differs from existing techniques in the following features. The same algorithm is used in tracking both moving and stationary objects by making the stationary objects "move." It is general enough to handle objects of various types and sizes, including point objects. The system has a fast implementation because all image operations are applied to small image regions only. We demonstrate the effectiveness of the proposed algorithm on a few real video sequences.
Standard image compression algorithms may not perform well in compressing images for pattern recognition applications, since they aim at retaining image fidelity in terms of perceptual quality rather than preserving spectrally significant information for pattern recognition. New compression algorithms for pattern recognition are therefore investigated, which are based on the modification of the standard compression algorithms to simultaneously achieve higher compression ratio and improved pattern recognition performance. This is done by emphasizing middle and high frequencies and discarding low frequencies according to a new distortion measure for compression. The operations of denoising, edge enhancement, and compression can be integrated in the same encoding process in the proposed compression algorithms. Simulation results show the effectiveness of the proposed compression algorithms.
We show that the standard image compression algorithms are not suitable for compressing images in correlation pattern recognition since they aim at retaining image fidelity in terms of perceptual quality rather than preserving spectrally significant information for pattern recognition. New compression algorithms for pattern recognition are therefore developed, which are based on the modification of the standard compression algorithms to achieve higher
compression ratio and simultaneously to enhance pattern recognition performance. This is done by emphasizing middle and high frequency components and discarding low frequency components according to a new developed distortion measure for compression. The operations of denoising, edge enhancement and compression can be integrated in the encoding process in the proposed compression algorithms. Simulation results show the effectiveness of the proposed compression algorithms.
A new filtering algorithm is presented which can remove impulse noise from corrupted images while preserving details. The algorithm is based on a new impulse detection technique that uses image gradients. The proposed impulse detector can effectively categorize all the pixels in an image into two classes -- noise pixels and noise-free pixels. The noise-free pixels are kept untouched while the noise pixels are filtered by a noise cancellor such as median filter. Experimental results show that the proposed algorithm provides significant improvement over many existing techniques in terms of both subjective and objective evaluations. It also has the advantage of computational simplicity over those algorithms.
Nonlinear morphological correlation provides some better performances in pattern recognition than linear correlation. However, it requires considerable amount of computational effort to obtain the final result. An adaptive threshold decomposition technique is proposed to reduce the computation and increase the processing speed. Computer simulation shows that the proposed method yields similar results to the original morphological correlation but with much less computational effort. A visual-area-coding technique is proposed to implement the morphological correlation optically in a single step. This alternative optical implementation provides several advantages over the optical morphological correlation schemes.
One-channel representations of hit-miss transform are described. The hit-miss transform can be reduced to be a cross-correlation followed by a thresholding operation after combining the foreground and background structuring elements into one structuring element. A phase-only joint transform correlator is proposed to realize one of the representations. Simulation results are provided.
We investigate the performance of morphological correlation under different illumination conditions. The morphological correlation is shown to be invariant to uniform input-image illumination when the input-image illumination is higher than that of the reference. Accordingly, illumination- independent pattern recognition can be realized using morphological correlation provided the reference is multiplied by a proper number less than unity or the input image is multiplied by a number greater than unity before threshold-decomposition. In addition, the correlation peak tends to get broader when illumination of the input is different from that of the reference. Computer simulation results are provided.
A new optoelectronic fuzzy inference system is proposed for processing a large number of fuzzy rules in parallel. The proposed system using spatial light modulator implements various membership functions as well as max-min inference. It has the features of easy implementation and large data processing capability. The membership function decomposition method in the improved fuzzy associative memory is used to save both space bandwidth and accommodate multiple-input fuzzy inference.
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.