A Novel Reduced-Layer Deep Learning System via Pixel Rearrangement for Object Detection in Multispectral Imagery

Abstract

In this paper, we study the problem of object detection on multispectral images. We present a generalized “Fully” Convolutional Neural Network (FCNN)-based deep learning system with a novel Pixel Rearrangement Technique, with reduced computational complexity and improved prediction accuracy than its state-of-the-art counterparts. In particular, we (a) define a key strategy based on spectral signatures to select a set of highly informative multispectral bands for the system; (b) for the first time, introduce a pixel rearrangement technique that efficiently utilizes pixels from the network's feature maps that results into accurate pixelwise prediction images; (c) propose dual stage global and adaptive thresholding methodologies that transform the pixelwise prediction images to binary. We evaluate the proposed system for automatic airborne building detection using the SpaceNet dataset. We use the three NVIDIA GeForce GTX 1060 GPUs at CMINDS Research Center and Tensorflow deep learning framework to implement the proposed system. Our findings show an improvement in the performance by 0.3% in comparison to the top winning submission of the national SpaceNet Building Challenge II, that took place in April 2017, but with an additional 43% reduction in the number of FCNN layers. Finally, we also present a comparison chart with various existing approaches to highlight the proposed reduced computational complexity system.