Method for Identifying Scene Graph Patterns associated with Image Classifier Predictions

A computer-implemented method of identifying patterns in scene graphs of images that correlate with correct or incorrect image classifications of a pre-trained image classifier, the method comprising:

• obtaining a scene graph for each of a plurality of images;
• classifying the plurality of images using the pre-trained image classifier;
• determining whether the predicted classification of each image is correct or incorrect by comparing the predicted classification with a corresponding expected classification;
• grouping the scene graphs into a first group and a second group based on whether the classification by the image classifier is correct or incorrect, respectively;
• and
• extracting, for each of the first group and the second group, a set of representative subgraphs shared among the scene graphs within the respective group, wherein the representative subgraphs represent the patterns correlated with the correct and incorrect classifications, respectively.