This thesis proposes a process for automatically validating that the LED display on a residential gateway reacts correctly to status changes. The process is based on object detection with Faster R-CNN and can potentially reduce both the workload of the tester and the risk of human error.
One of the main challenges is the small size of the LEDs on the image. To deal with this, a two-step method for small object detection is proposed that first detects the display as a region of interest and then detects the LED within that region of interest. This method is shown to perform well in comparison with other object detection models employing feature pyramid networks and similar techniques. The results indicate that the two-step approach shows best performance from 0.5 to 0.75 IoU levels. In other words, when speed is not a priority, image hierarchies can be more effective for small object detection than using feature pyramid network or similar techniques.
Additionally, the proposed method is shown to improve detection accuracy even when the original input image is the same size as the the input to the convolutional neural network, i.e. when cropping to the region of interest does not increase the number of significant pixels in the object to be detected.
Once detected, the behavior of each LED is categorized based on color and a comparison of subsequent frames in a video of the residential gateway. The behavior can then be compared to an expectation of what the behavior should be for a given gateway status.
The complete test process is demonstrated to work in lab conditions, and some improvements are proposed in order to generalize the testing. With these improvements, it is feasible to use the test process in day-to-day operation, thus reducing the cost of testing and providing a tangible benefit to the business. The work is expected to have limited impact on society since the scope of the work is very narrow and does not use any personal data.
2022. , p. 60