Distributed and self-validating computer vision for dense object detection in digital images

Abstract

A system for object recognition and segmentation from digital images provides an intelligent object recognition and segmentation using one or more multilayer convolutional neural network (CNN) models trained in multiple-stages and in a parallel and distributed manner to improve training speed and efficiency. The training dataset used in each of the multiple training stages for the CNN models are generated, expanded, self-validated from a preceding stage. The trained final CNN models are augmented with post-model filters to enhance prediction accuracy by removing false positive object recognition and segmentation. The system provides improved accuracy to predict object labels to append to unlabeled image blocks in digital images. In one embodiment, the system may be useful for enhancing a digital landmark registry by appending identifying labels on new infrastructure improvements recognized in aerial or satellite land images.