Vehicle damage validation using symmetry-based deep learning

A computer-implemented method for analyzing vehicle damage, comprising:

• receiving, at one or more processors, first and second images from opposing sides of a vehicle;
• detecting, using the one or more processors, a damage region in the first image;
• computing, by a neural network, deep learning feature vectors, wherein the deep learning feature vectors comprising:
• (a) learned visual representations of the damage region in the first image and a corresponding region in the second image,
• (b) numerical representations of visual patterns learned through neural network training to discriminate between damage patterns and normal vehicle features;
• computing a similarity measure between the deep learning feature vectors of the damage region and the corresponding region using a metric selected from one or more members of a group consisting of:
• a cosine similarity, an Euclidean distance, and specialized comparison algorithms;
• and
• validating the detected damage region based on the similarity measure;
• wherein the validating the detected damage region is performed by:
• computing an error measure between corresponding deep learning feature vectors from opposing sides of the vehicle to detect asymmetric visual patterns;
• filtering false positive detections caused by symmetric design features of the vehicle based on the error measure;
• and
• classifying the damage region as validated damage when the error measure exceeds a learned asymmetry threshold.