With the advent of the big data and intelligence eras, information systems require considerably enhanced performance and low energy costs. Optical computing may become the next-generation computing platform owing to its parallel processing capability and high bandwidth with low energy consumption. In pattern recognition applications, large amounts of image data must be rapidly processed. Two types of optical approaches have been investigated for pattern recognition: optical neural network, which comprises two subclasses including silicon photonic-based neural networks, and free-space-based optical network. The former has undergone considerable advancements recently owing to improved fabrication capability and novel network components based on optics such as microring resonators and Mach-Zehnder interferometers. The latter (e.g., diffractive neural networks) is also important, particularly for computational imaging-based applications. However, optical neural network-based pattern recognition approaches are immature owing to the implementation of nonlinear functions. Pattern recognition approaches founded on free-space-based optical networks are hybrid optoelectronic correlators, far more mature than optical neural network-based ones. The correlator can be codesigned with a neural network to serve as a coprocesser to prefilter some image features for ultrafast processing. However, in conventional optical correlators, both the spatial and spectral bandwidths of systems have not been efficiently used when performing the correlation operation. Hence, the inherent parallel processing capability of optics cannot be fully exploited.
In our previous work, a multichannel joint transform correlation method is proposed based on the compression and translation of joint transform power spectrum to fully utilize spatial and spectral bandwidths and enhance the parallel processing efficiency and recognition accuracy of optical correlation systems. In the input plane of this scheme, the scene image and N numbers of reference images are uploaded on different zones of the input spatial light modulator; then, the phase maps optimized using the iterative algorithm are superimposed onto the images. In the Fourier plane, interference between the Fourier spectra of scene images and those of every single reference image occurs in different zones of the Fourier plane. When the restriction parameter in the phase optimization algorithm is appropriated adjusted, no interference of the Fourier spectra of the reference images is observed. Consequently, the parallel processing of N channels is achieved without crosstalk. The relation between the localized peak clutter mean of the Fourier spectra of the preferred phase and the standard deviation of the correlation peak position is analyzed and used as a criterion for preferential preferred phase mask selection. Furthermore, the standard deviation of the correlation peak position is obtained for recognition tasks. In this study, we focus on distortion-invariant pattern recognition by integrating the multichannel joint transform correlator and the synthetic discriminant function. First, the feasibility of the local peak to clutter mean as a constraint for preferential phase selection is analyzed; results indicated that this factor is not appropriate when the synthetic discriminant function is used. Hence, a new phase selection criterion—known as the variation in the correlation peak position—is proposed to obtain the public preferred phase for targets with a specific distortion range. Furthermore, the selected phase is used in the multichannel joint transform correlator with the synthetic discriminant function to achieve distortion-invariant pattern recognition. Then, to determine the system performance in terms of the distortion level, the tolerance of our system on the scaling-down of the size of target and the increase in the number of training images for the synthetic discriminant function are analyzed. Finally, considering that the background may vary in real applications, we take successive video frames as varied input backgrounds and analyze the feasibility of our proposal.
Results indicate that under the considered image file size and background complexity, the proposed method can achieve nine-channel parallel recognition (Fig. 8). For correct recognition, the minimum scaling down factor is 0.6 (Fig. 11). When the number of rotated training images is increased to 9 in the synthetic discriminant function, a correct recognition can be guaranteed (Fig. 13). The relation between the minimum threshold of phase the optimization constraint and the synthesized image numbers of SDF is obtained for calculating the preferred phases (Table 1). Furthermore, a correct recognition can be guaranteed when the values of half the pixels in the background have changed (Fig. 16).
Herein, a novel distortion-invariant pattern recognition method based on multichannel joint transform correlator is proposed. The local peak to clutter mean is shown to be unsuitable, and we propose a new optimization criterion known as the variation in the correlation peak position, which is feasible in this proposal. We achieve nine-channel pattern recognition within 0.6-1.0 times of the scaling of the image size and rotation ranges of 0°-30°, 70°-100°, 140°-170°, 210°-240°, and 280°-310°. The upper limit of the number of synthesized training images is analyzed, which is nine in this study. Moreover, the proposed method can maintain its performance when the background is varied within the values of half its pixel, indicating robustness to background changes. The recognition speed and accuracy on distortion of the system are considerably improved with our proposal, which will benefit the development of practical multichannel optical correlators.
