Precision defect detection based on image difference with respect to templates
A computer-implemented method is provided for image-based defect detection. The method includes performing, by a processor device, template matching and subtraction on a set of training images and at least one template image to obtain a set of difference images. The difference images have defects, if any, highlighted therein. The method further includes generating, by the hardware processor applying a binary classification model to each of the training images in the set, activation heatmaps. The method also includes identifying, by the hardware processor, rough defect areas of interest in the activation heatmaps. The method additionally includes super-imposing, by the hardware processor, the activation heatmaps onto the difference images to obtain a set of super-imposed images, and highlight, as true defect areas, any areas in the super-imposed images having the defects from the difference images that overlap with the rough defect areas of interest from the activation heatmaps.
1. A computer-implemented method for image-based defect detection, comprising:
performing, by a processor device, template matching and subtraction on a set of training images and at least one template image to obtain a set of difference images, the difference images having defects, if any, highlighted therein;
generating, by the hardware processor applying a binary classification model to each of the training images in the set, activation heatmaps;
identifying, by the hardware processor, rough defect areas of interest in the activation heatmaps; and
super-imposing, by the hardware processor, the activation heatmaps onto the difference images to obtain a set of super-imposed images, and highlight, as true defect areas, any areas in the super-imposed images having the defects from the difference images that overlap with the rough defect areas of interest from the activation heatmaps.
2. The computer-implemented method of claim 1 , wherein the set of training images include class-labels and lack location labels for any of the defects therein.
3. The computer-implemented method of claim 1 , further comprising training a deep neural network for image segmentation using the set of training images.
4. The computer-implemented method of claim 3 , further comprising:
applying an inspection image of an object to the trained binary classification model in order to identify a defect area in the object; and
discarding the object responsive to identifying the defect area.
5. The computer-implemented method of claim 1 , further comprising training a deep neural network for object detection using the set of training images.
6. The computer-implemented method of claim 5 , further comprising:
applying an inspection image of an object to the trained binary classification model in order to identify the object; and
automatically engaging one or more motor vehicle functions to avoid contacting the object.
7. The computer-implemented method of claim 1 , wherein each of the super-imposed images consist of a respective single image having six channels, with three channels supplied from each of the constituent images forming the respective single image.
8. The computer-implemented method of claim 1 , further comprising screening out false defect areas in the difference images based on the heatmaps.
9. The computer-implemented method of claim 1 , wherein the false defect areas are detected in the difference images based on a lack of co-occurrence of a respective one of the rough defect areas of interest in a respective one of the heatmaps.
10. The computer-implemented method of claim 1 , further comprising ignoring, as false positives, any of the defects from the difference images that lack overlap with the rough defect areas of interest from the activation heatmaps.
11. A computer program product for image-based defect detection, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
performing, by a processor device, template matching and subtraction on a set of training images and at least one template image to obtain a set of difference images, the difference images having defects, if any, highlighted therein;
generating, by the hardware processor applying a binary classification model to each of the training images in the set, activation heatmaps;
identifying, by the hardware processor, rough defect areas of interest in the activation heatmaps; and
super-imposing, by the hardware processor, the activation heatmaps onto the difference images to obtain a set of super-imposed images, and highlight, as true defect areas, any areas in the super-imposed images having the defects from the difference images that overlap with the rough defect areas of interest from the activation heatmaps.
12. The computer program product of claim 11 , wherein the set of training images include class-labels and lack location labels for any of the defects therein.
13. The computer program product of claim 11 , wherein the method further comprises training a deep neural network for image segmentation using the set of training images.
14. The computer program product of claim 13 , wherein the method further comprises:
applying an inspection image of an object to the trained binary classification model in order to identify a defect area in the object; and
discarding the object responsive to identifying the defect area.
15. The computer program product of claim 11 , wherein the method further comprises training a deep neural network for object detection using the set of training images.
16. The computer program product of claim 15 , wherein the method further comprises:
applying an inspection image of an object to the trained binary classification model in order to identify the object; and
automatically engaging one or more motor vehicle functions to avoid contacting the object.
17. The computer program product of claim 11 , wherein each of the super-imposed images consist of a respective single image having six channels, with three channels supplied from each of the constituent images forming the respective single image.
18. The computer program product of claim 11 , wherein the method further comprises screening out false defect areas in the difference images based on the heatmaps.
19. The computer program product of claim 11 , wherein the false defect areas are detected in the difference images based on a lack of co-occurrence of a respective one of the rough defect areas of interest in a respective one of the heatmaps.
20. A computer processing system for image-based defect detection, comprising:
a memory for storing program code; and
a processor device for running the program code to
perform template matching and subtraction on a set of training images and at least one template image to obtain a set of difference images, the difference images having defects, if any, highlighted therein;
generate, by applying a binary classification model to each of the training images in the set, activation heatmaps;
identify rough defect areas of interest in the activation heatmaps; and
super-imposing the activation heatmaps onto the difference images to obtain a set of super-imposed images, and highlight, as true defect areas, any areas in the super-imposed images having the defects from the difference images that overlap with the rough defect areas of interest from the activation heatmaps.