Joseph P. Garcia

Wavelet transforms and neural networks for compression and recognition

Abstract Robust recognition for image and speech processing needs data compression that preserves features. To accomplish this, we have utilized the discrete wavelet transforms and the continuous wavelet transforms (CWT) together with artificial neural networks (ANN) to achieve automatic pattern recognition. Our approach is motivated by the mathematical analog of the CWT to the human hearing and visual systems, e.g., the so-called Mexican hat and Gabor functions, Gaussian window, respectively. We develop an ANN method to construct an optimum mother wavelet that can organize sensor input data in the multiresolution format that seems to become essential for brainstyle computing. In one realization, the architecture of our ANN is similar to that of a radial basis function approach, except that each node is a wavelet having three learnable parameters: weight W ij , scale a, and shift b. The node is not a McCullouch-Pitts neuron but a “wave-on”. We still use a supervised learning conjugate gradient descent algorithm in these parameters to construct a “super-mother” wavelet from a superposition of a set of waveons-mother wavelets. Using these techniques, we can accomplish the signal-enhanced and feature-preserving compression, e.g., on the infrared images, that avoids the overtraining and overfitting that have plagued ANNs ability to generalize and abstract information.

Adaptive wavelet classification of acoustic backscatter and imagery

The utility and robustness of wavelet features is demonstrated through three practical case studies of detecting objects in multispectral electro-optical imagery, sidescan sonar imagery, and acoustic backscatter. Attention is given to choosing proper waveforms for particular applications. Using artificial neural networks (ANNs), evidence is fused from multiple-waveform types that detect local features. The wavelet waveforms and their dilation and shift parameters are adaptively computed with ANNs to maximize classification accuracy. Emphasis is placed on the acoustic backscatter case study, involving detecting a metallic man-made object from natural and synthetic specular clutter with reverberation noise. The synthetic clutter is shown to be a good model for the natural clutter, which appears promising for avoiding huge data collection efforts for natural clutter and for better delineating the classification boundary. The classifier computes the locations, sizes, and weights of Gaussian patches in time-scale space that contain the most discriminatory information. This new approach is shown to give higher classification rates than an ANN with commonly used power spectral features. The new approach also reduces the number of free parameters in the classifier based on all wavelet features, which leads to simpler implementation for applications and to potentially better generalization to test data.

Super-Haar designs of wavelet transforms

A linear superposition of Haar transform is given to design an adaptive biorthogonal subband coding. Given Haar scaling function, rect(x), a new symmetric staircase scaling function may be desirable in order to match a specific compression or recognition goal. Then, the associated lowpass filter coefficients are solved from the roots on a unit circle in the z-transform domain. From which the bi-orthogonal and lossless high pass filter is derived for both the forward analysis and the backward synthesis stages. An explicit construction of super-Haar system produces a lossless filter bank which can match an infrared image for achieving pattern recognition and compression simultaneously.

Subpixel jitter video restoration on board of micro-UAV

We review various image processing algorithms for micro-UAV EO/IR sub-pixel jitter restoration. Since the micro-UAV, Silver Fox, cannot afford isolation coupling mounting from the turbulent aerodynamics of the airframe, we explore smart real-time software to mitigate the sub-pixel jitter effect. We define jitter to be sub-pixel or small-amplitude vibrations up to one pixel, as opposed to motion blur over several pixels for which there already exists real time correction algorithms used on other platforms. We divide the set of jitter correction algorithms into several categories: They are real time, pseudo-real time, or non-real-time, but they are all standalone, i.e. without relying on a library storage or flight data basis on-board the UAV. The top of the list is demonstrated and reported here using real-world data and a truly unsupervised, real-time algorithm.

Fingerprint data acquisition, desmearing, wavelet feature extraction, and identification

In this paper, we present (1) a design concept of a fingerprint scanning system that can reject severely blurred inputs for retakes and then de-smear those less blurred prints. The de-smear algorithm is new and is based on the digital filter theory of the lossless QMF (quadrature mirror filter) subband coding. Then, we present (2) a new fingerprint minutia feature extraction methodology which uses a 2D STAR mother wavelet that can efficiently locate the fork feature anywhere on the fingerprints in parallel and is independent of its scale, shift, and rotation. Such a combined system can achieve high data compression to send through a binary facsimile machine that when combined with a tabletop computer can achieve the automatic finger identification systems (AFIS) using todays technology in the office environment. An interim recommendation for the National Crime Information Center is given about how to reduce the crime rate by an upgrade of todays police office technology in the light of the military expertise in ATR.

Composite wavelet features for image recognition

A detection technique based on a synergistic composition of wavelet feature detectors is demonstrated on sonar imaging data. The wavelets are used to preprocess the imagery for enhancing highlights and shadows. A neural network is trained on the preprocessed imagery to weight the output of two filters for underwater object detection. This approach is demonstrated on multiple scales. Results indicate this composite approach is highly effective.

Molecular computing for edge-enhanced laser imaging

In order to illustrate the self-assembly capability, we consider a laser imaging experiment on a wet film that is made of bacteriorhodopsin (BR) molecules suspended in a diffusion-limited viscous medium. BR wet film is similar to a wet photograph film but having a finer resolution and adaptive pixel locations due to laser-induced thermal diffusion. The synergism between thermal diffusion of BR molecules (induced externally by a write-laser) and molecular photochromism (generated internally by a read-laser) is exploited naturally for edge-enhanced image applications.

Chords in wavelet projection transform space applied to aspect invariant pattern recognition

We describe a local projection transform that uses the discrete and continuous wavelet transforms to represent edge features over multiple scales. High order aspect invariant features are generated from chords in the transform space. A novel noise coding technique binds these features into a coherent pattern. This approach permits a reduction of the dimensionality of the information giving rise to lower processing requirements and thus permitting implementation of automatic target recognition (ATR) on conventional computer architectures.

Sensor fusion for wide-area surveillance

The Gabor transform (GT) is applied to the super-resolution of noisy dot image on the infrared focal plane array (FPA) for the remote surveillance of aircraft or missiles. A unique solution of this kind of ill-posed problem is possible because we have incorporated the measured or a priori known size information of the engine/nozzle. Yet noise makes the super-resolution ill- conditioned. We surmount this difficulty by incorporating the GT into a modified Papoulis- Gerchberg iteration algorithm. This is possible because the GT is a local Fourier transform (FT), it matches the localized object signal (object size one unit) but mismatches the global nature of noise. In a practical case of a photon-limited signal having a signal to noise ratio as low as 1.3, our approach recognizes a simulated missile plume. We also show additional resolution can be gained if the radar backscattering from the nozzle and other scatterers is fused with the spatially resolved single image pixel.

Integrated image compression and detection for minelike objects

The need is described for a system-level integrated treatment of compression and detection methods and several issues are raised. Compression detection examples are provided as a first step in this direction and to illustrate the concepts.