Performance Assessment of the Logarithmic-hybrid Optical Neural Network Filter for Multiple Objects Recognition

摘要

Previously, we have described the complex logarithmic r-theta mapping for hybrid optical neural network (L-HONN) filter for object recognition within cluttered scenes. The design of the filter is based on a space-variant imaging sensor combined with the hybrid optical neural network filter. The created L-HONN filter exhibits simultaneously and with a single pass over the input data in-plane rotation, out-of-plane rotation, scale and projection, and shift invariance. The window unit in the design of the L-HONN filter allows multiple objects of the same class to be detected and classified within the input image. Additionally, the architecture of the neural network unit of the filter allows the recognition of multiple objects of different classes within the input image by augmenting the output layer of the unit. We have presented results for multiple objects recognition within cluttered scenes. We have tested the filter with both still images and video sequences. L-HONN filter is able to successfully suppress the unknown background clutter and recognize the different classes of the objects in the input scene of a still image or a video frame. As for all the HONN-type filters, it is found that by altering the target classification levels for each of the object classes, the L-HONN filter's behaviour can be varied to suit different application requirements from more like a high-pass biased filter to more like a minimum variance synthetic discriminant function (MVSDF) filter. Here, for first time, we present initial results of the full-series of tests for examining the performance of the L-HONN filter for its detectability and peak sharpness, performance with in-class distortion of the input objects, its discrimination ability between in-class and out-of-class objects and we record the results for the filter's tolerance-to-clutter in recognising the objects of different classes. From the performance analysis, L-HONN filter is shown to exhibit good peak correlation energy (PCE) values, good distortion tolerance by recognising the true-class object out-of-plane rotated for a range of 10° to 90° degrees, and it is able to discriminate between the different classes objects. Finally, it is shown to successfully recognise the areas where the true-class objects are located.