Amphora detection based on a gradient weighted error in a convolution neuronal network

Abstract

In this paper, we propose a method based on pixel prediction to detect objects into a large image. We propose to integrate the Weighted Error Layer (WEL) in a Convolution Neuronal Network (CNN) architecture in order to weight the error during the back-propagation and to reduce the impact of the borders. We estimate the orientation of the objects when the detection step is achieved. Our proposed layer is evaluated on real data in order to detect amphorae on the Mazatos underwater archaeological site. © (2017) by the International Measurement Confederation (IMEKO). All rights reserved.