Machine learning-based image recognition of weather damage
Various image analysis techniques are disclosed herein that automatically assess the damage to a rooftop of a building or other object. In some aspects, the system may determine the extent of the damage, as well as the type of damage. Further aspects provide for the automatic detection of the roof type, roof geometry, shingle or tile count, or other features that can be extracted from images of the rooftop.
1. A method comprising:
receiving, at a device, an image of a rooftop;
dividing, by the device, the image into a plurality of subdivision images, based on one or more characteristics of rooftop shingles or tiles depicted in the subdivision images by subdividing the image of the rooftop such that each of a plurality of the subdivision images depicts a) a number of shingle tabs or tiles within a predefined count range and b) damage to the shingles or tiles in a given subdivision image centered in the subdivision image, wherein the one or more characteristics comprise i) a count of shingle tabs or tiles and ii) damage to the shingles or tiles;
applying, by the device, a machine learning-based classifier comprising a convolutional neural network to one or more of the subdivision images, wherein the classifier is configured to detect damage done to a particular one of the shingles or tiles depicted in a particular subdivision image under analysis and to assign a damage type to the detected damage; and
sending, by the device, display data for display that is indicative of an extent of damage to the rooftop associated with the assigned damage type.
2. The method as in claim 1 , wherein the damage type is indicative of at least one of: hail damage, ice damage, wind damage, wear and tear damage, or human-made damage.
3. The method as in claim 1 , further comprising:
performing edge detection on the rooftop shingles or tiles in the image, to identify individual shingle tabs or tiles; and
determining the count of the individual shingle tabs or tiles depicted in each of the plurality of subdivision images.
4. The method as in claim 1 , further comprising:
estimating, by the device, a shingle tab or tile count for the rooftop, based on determined counts of individual shingle tabs or tiles in the image; and
estimating, by the device, an area of damaged shingle or tiles of the rooftop associated with the damage type.
5. The method as in claim 1 , wherein subdividing the image of the rooftop such that each of a plurality of the subdivision images depicts a number of shingle tabs or tiles within a predefined count range comprises:
applying, by the device, a second machine learning-based classifier to portions of the image of the rooftop, wherein the second classifier was trained to output a shingle tab or tile count for an assessed portion of the image of the rooftop.
6. The method as in claim 1 , further comprising:
forming, by the device, a three dimensional model of the rooftop, in part by correlating points or edges depicted in a plurality of images of the rooftop.
7. The method as in claim 1 , further comprising:
applying, by the device, a second machine learning-based classifier to the image of the rooftop, to identify a roof type associated with the rooftop comprising at least one of: asphalt shingled, fiberglass shingled, wood shingled, tile, slate, metal, or rubber membrane.
8. The method as in claim 1 , wherein the image of the rooftop was captured by an unmanned aerial vehicle (UAV).
9. The method as in claim 1 , further comprising:
applying, by the device, a second machine learning-based classifier to the image of the rooftop, to detect a boundary of the rooftop, wherein the classifier classifies the boundary as being one of: a ridge, a ridge edge, an occluding ridge, a hip, a hip edge, an occluding hip, a valley, a pitch change, an inside edge, an outside edge, or an occluding edge.
10. The method as in claim 1 , further comprising:
generating, by the device, a pseudo-orthomosaic image that combines the image of the rooftop with a plurality of other images of the rooftop by:
projecting the images of the rooftop onto a ground to form a Voroni map;
mapping the images of the rooftop to their corresponding Voroni regions of the map; and
sending, by the device, the pseudo-orthomosaic image for display.
11. The method as in claim 1 , wherein the machine learning-based classifier comprises a multilayer perceptron (MLP) model.
12. The method as in claim 1 , further comprising:
identifying, by the device, individual slopes of the rooftop based in part on the image of the rooftop.
13. The method as in claim 12 , further comprising:
performing, by the device, damage detection on one or more subdivision images on a per slope basis.
14. An apparatus, comprising:
one or more network interfaces to communicate with a network;
a processor coupled to the network interfaces and configured to execute one or more processes; and
a memory configured to store a process executable by the processor, the process when executed configured to:
receive an image of a rooftop;
divide the image into a plurality of subdivision images, based on one or more characteristics of rooftop shingles or tiles depicted in the subdivision images by subdividing the image of the rooftop such that each of a plurality of the subdivision images depicts a) a number of shingle tabs or tiles within a predefined count range and b) damage to the shingles or tiles in a given subdivision image centered in the subdivision image, wherein the one or more characteristics comprise i) a count of shingle tabs or tiles and ii) damage to the shingles or tiles;
apply a machine learning-based classifier comprising a convolutional neural network to one or more of the subdivision images, wherein the classifier is configured to detect damage done to a particular one of the shingles or tiles depicted in a particular subdivision image under analysis and to assign a damage type to the detected damage; and
send display data for display that is indicative of an extent of damage to the rooftop associated with the assigned damage type.
15. The apparatus as in claim 14 , wherein the damage type is indicative of at least one of: hail damage, ice damage, wind damage, wear and tear damage, or human-made damage.
16. The apparatus as in claim 14 , wherein the process when executed is further configured to:
estimating, by the apparatus, a shingle tab or tile count for the rooftop, based on the determined counts of individual shingle tabs or tiles in the image; and
estimating, by the apparatus, an area of damaged shingle or tiles of the rooftop associated with the damage type.
17. A tangible, non-transitory, computer-readable medium storing program instructions that cause a device to execute a process comprising:
receiving, at the device, an image of a rooftop;
dividing, by the device, the image into a plurality of subdivision images, based on one or more characteristics of rooftop shingles or tiles depicted in the subdivision images by subdividing the image of the rooftop such that each of a plurality of the subdivision images depicts a) a number of shingle tabs or tiles within a predefined count range and b) damage to the shingles or tiles in a given subdivision image centered in the subdivision image, wherein the one or more characteristics comprise i) a count of shingle tabs or tiles and ii) damage to the shingles or tiles;
applying, by the device, a machine learning-based classifier comprising a convolutional neural network to one or more of the subdivision images, wherein the classifier is configured to detect damage done to a particular one of the shingles or tiles depicted in a particular subdivision image under analysis and to assign a damage type to the detected damage; and
sending, by the device, display data for display that is indicative of an extent of damage to the rooftop associated with the assigned damage type.
18. The tangible, non-transitory, computer-readable medium of claim 17 , wherein the damage type is indicative of at least one of: hail damage, ice damage, wind damage, wear and tear damage, or human-made damage.
19. The tangible, non-transitory, computer-readable medium of claim 17 , the process further comprising:
estimating a shingle tab or tile count for the rooftop, based on determined counts of individual shingle tabs or tiles in the image; and
estimating an area of damaged shingle or tiles of the rooftop associated with the damage type.
20. The tangible, non-transitory, computer-readable medium of claim 17 , the process further comprising:
applying a second machine learning-based classifier to the image of the rooftop, to detect a boundary of the rooftop, wherein the classifier classifies the boundary as being one of: a ridge, a ridge edge, an occluding ridge, a hip, a hip edge, an occluding hip, a valley, a pitch change, an inside edge, an outside edge, or an occluding edge.