Ioannis Kypraios

Object recognition within cluttered scenes employing a hybrid optical neural network filter

We propose a hybrid filter, which we call the hybrid optical neural network (HONN) filter. This filter combines the optical implementation and shift invariance of correlator-type filters with the nonlinear superposition capabilities of artificial neural network methods. The filter demonstrates good performance in maintaining high-quality correlation responses and resistance to clutter to nontraining in-class images at orientations intermediate to the training set poses. We present the design and implementation of the HONN filter architecture and assess its object recognition performance in clutter.

Performance assessment of the Modified-Hybrid Optical Neural Network filter

We present in detail the recorded results of the modified-hybrid optical neural network (M-HONN) filter during a full series of tests to examine its robustness and overall performance for object recognition tasks. We test the M-HONN filter for its detectability and peak sharpness with within-class distortion of the input object, its discrimination ability between an in-class and out-of-class object, and its performance with cluttered images of the true-class object. The M-HONN filter is found to exhibit good detectability, an ability to maintain its correlation-peak sharpness throughout the recorded tests, good discrimination ability, and an ability to detect the true-class object within cluttered input images. Additionally we observe the M-HONN filters performance within the tests in comparison with the constrained-hybrid optical neural network filter for the first three series of tests and the synthetic discriminant function-maximum average correlation height filter for the fourth set of tests.

Fully invariant object recognition in cluttered scenes

A method of detecting target objects in cluttered scenes despite any kind of geometrical distortion is demonstrated. Several existing techniques are combined, each one capable of creating invariance to one or more types of distortion of the target object. A MACH filter combined with an SDF creates invariance to orientation while constraining the correlation peak amplitudes and giving good tolerance to background clutter and noise. A log r-θ mapping is employed to give invariance to in-plane rotation and scale.

A nonlinear training set superposition filter derived by neural network training methods for implementation in a shift-invariant optical correlator

The various types of synthetic discriminant function (sdf) filter result in a weighted linear superposition of the training set images. Neural network training procedures result in a non-linear superposition of the training set images or, effectively, a feature extraction process, which leads to better interpolation properties than achievable with the sdf filter. However, generally, shift invariance is lost since a data dependant non-linear weighting function is incorporated in the input data window. As a compromise, we train a non-linear superposition filter via neural network methods with the constraint of a linear input to allow for shift invariance. The filter can then be used in a frequency domain based optical correlator. Simulation results are presented that demonstrate the improved training set interpolation achieved by the non-linear filter as compared to a linear superposition filter.

Performance assessment of Unconstrained Hybrid Optical Neural Network (U-HONN) filter for object recognition tasks in clutter

Previously we have described a hybrid optical neural network (HONN) filter. The filter is synthesised employing an artificial neural network technique that generates a non-linear interpolation of the intermediate train set poses of the training-set objects but maintains linear shift-invariance which allows potential implementation within a linear optical correlator type architecture. In this paper, we remove the constraints imposed on the filter’s output correlation peak height from the constraint matrix of the synthetic discriminant function used to create the composite filter. We examine the U-HONN filter’s detectability, peak sharpness, within-class distortion range, discrimination ability between an in-class and out-of-class object and the filter’s tolerance to clutter. We assess the behaviour of the U-HONN filter in an open area surveillance application. The filter demonstrates good object detection abilities within cluttered scenes, keeping good quality correlation peak sharpness and detectability throughout all the sets of tests. Thus the U-HONN filter is able to detect and accurately classify the in-class object within different background scenes at intermediate angles to the train-set poses.

Implementation of the Maximum Average Correlation Height (MACH) filter in the spatial domain for object recognition from clutter backgrounds

A moving space domain window is used to implement a Maximum Average Correlation Height (MACH) filter which can be locally modified depending upon its position in the input frame. This enables adaptation of the filter dependant on locally variant background clutter conditions and also enables the normalization of the filter energy levels at each step. Thus the spatial domain implementation of the MACH filter offers an advantage over its frequency domain implementation as shift invariance is not imposed upon it. The only drawback of the spatial domain implementation of the MACH filter is the amount of computational resource required for a fast implementation. Recently an optical correlator using a scanning holographic memory has been proposed by Birch et al [1] for the real-time implementation of space variant filters of this type. In this paper we describe the discrimination abilities against background clutter and tolerance to in-plane rotation, out of plane rotation and changes in scale of a MACH correlation filter implemented in the spatial domain.

On a method to eliminate moving shadows in video sequences

We present a simple computational model that works in the RGB colour space to detect moving shadow pixels in video sequences of indoor scenes, illuminated in each case by an incandescent source. A channel ratio test for shadows cast on some common indoor surfaces is proposed that can be appended to the developed scheme so as to reduce the otherwise high false detection rate. The core method, based on a Lambertian hypothesis, has been adapted to work well for near-matte surfaces by suppressing highlights. The results reported, based on an extensive data analysis conducted on some of the crucial parameters involved in the model, not only bring out the subtle details of the parameters, but also remove the ad hoc nature of the chosen thresholds to a certain extent. The method has been tested on various indoor video sequences; the results obtained indicate that it can be satisfactorily used to mark or eliminate the strong portion of the foreground shadow region.

Performance Analysis of a Modified Moving Shadow Elimination Method Developed for Indoor Scene Activity Tracking

Moving shadow detection is an important step in automated robust surveillance systems in which a dynamic object is to be segmented and tracked. Rejection of the shadow region significantly reduces the erroneous tracking of non-target objects within the scene. A method to eliminate such shadows in indoor video sequences has been developed by the authors. The objective has been met through the use of a pixel-wise shadow search process that utilizes a computational model in the RGB colour space to demarcate the moving shadow regions from the background scene and the foreground objects. However, it has been observed that the robustness and efficiency of the method can be significantly enhanced through the deployment of a binary-mask based shadow search process. This, in turn, calls for the use of a prior foreground object segmentation technique. The authors have also automated a standard foreground object segmentation technique through the deployment of some popular statistical outlier-detection based strategies. The paper analyses the performance i.e. the effectiveness as a shadow detector, discrimination potential, and the processing time of the modified moving shadow elimination method on the basis of some standard evaluation metrics.

Logarithmic r-θ Map for Hybrid Optical Neural Network Filter for Object Recognition within Cluttered Scenes

We combine the complex logarithmic r-thetas mapping of a space-variant imaging sensor with the hybrid optical neural network filter for achieving an overall out-of-plane rotation, in-plane rotation, scale and projection invariance and is resistance to clutter. The resulted filter is called the complex logarithmic r-thetas mapping for the hybrid optical neural network (L-HONN) filter. For restoring the shift invariance of the input images of the objects, lost by applying to the images the logarithmic mapping, we include in the filterpsilas design a window-based unit. We assess the performance and record the results of the L-HONN filter with cluttered object images.

Image watermarking extraction using Fourier domain Wiener filter

Digital watermarking is a vital process for protecting the copyright of images. This paper presents a method of embedding a private robust watermark into a digital image. The full complex form the Wiener filter is used to extract the signal from the watermarked image. This is shown to outperform the more conventional approximate notation. The results are shown to be extremely noise insensitive.