Broad area geospatial object detection using autogenerated deep learning models
View Patent ↗A system for automated geospatial image analysis comprising a deep learning model module and a convolutional neural network serving as an automated image analysis software module. The deep learning module receives a plurality of orthorectified geospatial images, pre-labeled to demarcate objects of interest, and optimized for the purpose of training the neural network of the image analysis software module. The module presents marked geospatial images and a second set of unmarked, optimized, training geospatial images to the convolutional neural network. This process may be repeated so that an image analysis software module can detect multiple object types or categories. The image analysis software module receives a plurality of orthorectified geospatial images from one or more geospatial image caches. Using multi-scale sliding window submodule, image analysis modules scan geospatial images, detect objects present and locate them on the geographical latitude-longitude system. The system reports the results in the requestor's preferred format.
1. A system for broad area geospatial object detection using auto-generated deep learning models, comprising:
a deep learning model training software module stored in a memory of and operating on a processor of a computing device; and
an image analysis software module stored in the memory of and operating on the processor of the computing device;
wherein the deep learning model training software module:
(a) receives training data comprising a plurality of orthorectified geospatial images with a plurality of objects present therein, at least a first subset of the plurality of objects being labeled and a second subset of objects being unlabeled;
(b) aggregates the training data and classifies the training data into a plurality of predefined categories;
(c) applies one or more image modification steps to the training data, drawn from a set comprising artifact removal, color standardization, and resolution data determination;
(d) optimizes the training data for deep learning model training;
(e) generates an object classification model from the training data using a deep learning method comprising separate processing of the first and second subsets of the training data through a convolutional neural network system; and
wherein the image analysis software module:
(f) receives unanalyzed, orthorectified geospatial imagery;
(g) applies one or more image modification steps to the unanalyzed orthorectified geospatial imagery, drawn from a set comprising artifact removal, color standardization, and resolution data determination;
(h) optimizes the orthorectified geospatial imagery for object classification;
(i) discards images unsuitable for analysis;
(j) uses the object classification model generated in (f) to automatically identify and label all objects of interest in the received, unanalyzed orthorectified geospatial imagery, regardless of the orientation of that feature item within the section and accounting for differences in item scale by using a multi-scale sliding window algorithm; and
(k) outputs the locations of the identified objects of interest in a form dictated by the parameters of the original search request.
2. A method for conducting automated identification of feature items from orthorectified geospatial imagery data comprising the following steps:
(a) receiving a plurality of orthorectified geospatial imagery, divided into discrete segments where a plurality of objects of interest are present therein, at least a first subset of the plurality of objects of interest being labeled and a second subset of objects of interest being unlabeled, to a deep learning model training software module stored in a memory of and operating on a processor of a computing device;
b) placing pre-labeled geospatial imagery segments into pre-defined categories based upon the objects of interest they contain;
(c) applying one or more image modification steps drawn from a set comprising artifact removal, color standardization, and resolution data determination;
(d) optimizing the images for use in the deep learning model;
(e) discarding image segments deemed unsuitable for desired purpose;
(f) generating an object classification model from the training data using a deep learning method comprising separate processing of the first and second subsets of the training data through a convolutional neural network system operating as part of an image analysis software module stored in a memory of and operating on a processor of a computing device;
(g) using the image analysis software module trained as in steps (a) through (f) to identify feature items of interest within unanalyzed geospatial imagery data, regardless of the feature items' orientation and accounting for differences in item scale by using a multi-scale sliding window algorithm; and
(h) reporting the identities and locations of all feature items of interest in a form dictated by the parameters of the original search request.