CUD a-accelerated fast training of Locally connected Neural Pyramid using YIQ color coding

To achieve a panoptic machine vision, recognition of images from disparate classes like person, car, building, etcetera is of primal importance. The Locally-connected Neural Pyramid (LCNP) was proposed earlier to achieve a robust and a time efficient training of large datasets of images from these disparate classes. The objective of this paper is to propose a technique for fast training of the LCNP. YIQ coding is used to extract the color based information of the images as it separates the color information from the luminance information. As R GB to YIQ conversion is an embarrassingly parallel situation, this recoding can give a tremendous speed-up over the previous approach- where PCA de-correlation of RGB channels was carried out. Also, the use of YIQ coding has entailed a reduction in the complexity of the LCNP, thus, reducing the computations considerably. This will further boost the time performance of the training. Despite a considerable reduction in the complexity of the LCNP and the use of YIQ coding, the recognition performance achieved by this approach is similar to the previous approach. A recognition rate of 85.62% is achieved for the testing samples of the LabelMe-12-50K dataset. We p ropose that if the previous method of de-correlating RGB ch annels using PCA is rep laced with YIQ coding, tremendous speed-up will be achieved.