Automated visual inspection of syringes

1. A computer-implemented method for identifying whether a syringe is defective, the method comprising:

receiving an image of the syringe, the image comprising a foreground and a background, wherein a plurality of frames of the syringe that includes the image are received, each frame being a sequential image of the syringe over a span of time;

generating an updated image that accentuates the foreground by subtracting the background from the image;

applying a bounding box to a group of neighboring pixels in the updated image;

inputting the bounding box into a classifier;

receiving, as output from the classifier, a label indicating whether the syringe is defective;

responsive to the label indicating that the syringe is defective, tracking a trajectory of an object in the syringe across the plurality of frames; and

evaluating an accuracy of the output from the classifier based on the trajectory.

2. The computer-implemented method of claim 1 , wherein evaluating the accuracy of the output from the classifier comprises:

determining that the trajectory of the object is a downward trajectory; and

determining, based on the downward trajectory, that the object is a defect.

3. The computer-implemented method of claim 1 , wherein evaluating the accuracy of the output from the classifier comprises:

responsive to determining that the object is stationary across the plurality of frames, determining that the object is not a defect; and

responsive to determining that the object is not a defect, modifying the label output from the classifier, the modified label indicating that the syringe is not defective.

4. The computer-implemented method of claim 1 , wherein tracking the trajectory of the object comprises:

tracking movement of the object within a fluid in the syringe across the plurality of frames; and

determining the trajectory of the object from tracked movement of the object below a stopper of the syringe and greater than a threshold distance from a meniscus of fluid in the syringe.

5. The computer-implemented method of claim 4 , further comprising:

determining that the trajectory of the object is an upward trajectory;

responsive to determining that the trajectory is an upward trajectory, determining that the object is a bubble within the fluid in the syringe; and

modifying the label output from the classifier, the modified label indicating that the syringe is not defective.

6. The computer-implemented method of claim 1 , wherein subtracting the background from the image comprises:

identifying a static object, the static object stationary across the plurality of frames; and

removing the static object from the image of the syringe.

7. The computer-implemented method of claim 6 , wherein identifying the static object comprises:

determining a distribution of pixel intensities across the plurality of frames; and

identifying pixels with unchanged pixel intensities across the plurality of frames to identify the static object.

8. The computer-implemented method of claim 1 , wherein the classifier is a deep convolutional neural network.

9. The computer-implemented method of claim 1 , further comprising:

inputting the image of the syringe into the classifier; and

receiving, as output from the classifier, a label indicating whether the syringe is defective.

10. The computer-implemented method of claim 1 , further comprising:

inputting the updated image into the classifier; and

receiving, as output from the classifier, a label indicating whether the syringe is defective.

11. The computer-implemented method of claim 1 , wherein the label received as output from the classifier indicates a type of defect in the syringe.

12. The computer-implemented method of claim 11 , wherein the type of defect is one of fiber and dust particles.

13. The computer-implemented method of claim 1 , wherein neighboring pixels comprise pixels of a threshold intensity that are a threshold number of pixels away from each other.

14. The computer-implemented method of claim 1 , further comprising:

applying a second bounding box to a second group of neighboring pixels in the updated image;

inputting the second bounding box into the classifier; and

receiving as output from the classifier, a label indicating whether the syringe is defective.

15. The computer-implemented method of claim 14 , further comprising:

receiving as output from the classifier, a label indicating a type of defect associated with each of the plurality of bounding boxes.

16. A non-transitory computer readable storage medium comprising computer executable code that when executed by one or more processors causes the one or more processors to perform operations comprising:

receiving an image of the syringe, the image comprising a foreground and a background, wherein a plurality of frames of the syringe that includes the image are received, each frame being a sequential image of the syringe over a span of time;

generating an updated image that accentuates the foreground by subtracting the background from the image;

applying a bounding box to a group of neighboring pixels in the updated image;

inputting the bounding box into a classifier;

receiving, as output from the classifier, a label indicating whether the syringe is defective;

responsive to the label indicating that the syringe is defective, tracking a trajectory of an object in the syringe across the plurality of frames; and

evaluating an accuracy of the output from the classifier based on the trajectory.

17. The non-transitory computer-readable medium of claim 16 , wherein evaluating the accuracy of the output from the classifier comprises:

determining that the trajectory of the object is a downward trajectory; and

determining, based on the downward trajectory, that the object is a defect.

18. The non-transitory computer-readable medium of claim 16 , wherein evaluating the accuracy of the output from the classifier comprises:

responsive to determining that the object is stationary across the plurality of frames, determining that the object is not a defect; and

responsive to determining that the object is not a defect, modifying the label output from the classifier, the modified label indicating that the syringe is not defective.

19. The non-transitory computer-readable medium of claim 16 , wherein tracking the trajectory of the object comprises:

tracking movement of the object within a fluid in the syringe across the plurality of frames; and

determining the trajectory of the object from tracked movement of the object below a stopper of the syringe and greater than a threshold distance from a meniscus of fluid in the syringe.

20. A system comprising:

one or more computer processors; and

a non-transitory computer readable storage medium comprising computer executable code that when executed by the one or more processors causes the one or more processors to perform operations comprising:

receiving an image of the syringe, the image comprising a foreground and a background, wherein a plurality of frames of the syringe that includes the image are received, each frame being a sequential image of the syringe over a span of time;

generating an updated image that accentuates the foreground by subtracting the background from the image;

applying a bounding box to a group of neighboring pixels in the updated image;

inputting the bounding box into a classifier;

receiving, as output from the classifier, a label indicating whether the syringe is defective;

responsive to the label indicating that the syringe is defective, tracking a trajectory of an object in the syringe across the plurality of frames; and

evaluating an accuracy of the output from the classifier based on the trajectory.