Top 3 automatically detected image errors

First, the key data: In 7.5% of a person's photo uploads, the photo is automatically rejected. Three quarters of these people then submit an accepted photo on the second attempt, 17% need a third and 9% a fourth attempt.

On the fourth attempt, the AI is skipped and the image is accepted in any case. In this way, it can be ruled out that physically impaired persons are denied the upload by the AI. If the photo quality is insufficient, the photo can still be rejected during manual quality control.

The total number of successful photo uploads, on the other hand, is only marginally reduced by the first gate checks: 99.4% of all people manage to submit a photo accepted by the AI.

Reasons for rejection

Rank 1: At just under 42%, edge distance is the most common reason for rejection. This is rated as insufficient in a total of 3% of uploads. If the distance to the edge is too small, the image cannot be cropped nicely and gaps appear, mostly below the chin or due to cropped hairstyles. This problem often occurs with existing passport photos. Photo Collect defines a "Safety Zone" around the face, which must be filled with image content in any case.

Rank 2: A face with too low a resolution is objected to in 21% of the rejections or 1.6% of all uploads. Photo Collect looks not only at the resolution of the submitted photo, but also at the number of pixels in the relevant face area. According to the ICAO standard, there must be at least 100 pixels of space horizontally between the pupils. This error occurs if the resolution of the images is too small from the start (which is often the case with old image files), or if the person in the image is taken from too far away.

Rank 3: An incorrect head position is the trigger in 19% of rejections and occurs in 1.4% of all uploads. Here Photo Collect distinguishes between the frequently occurring left/right looking away ("yaw", 85%) and the shot from the frog/bird perspective ("pitch", 15%). The rotation of the head is automatically corrected by Photo Collect and therefore does not appear in these statistics.

Interesting: Pictures with hygiene masks were widespread in 2021 - from the middle of 2022, we hardly find such uploads. Group pictures with more than one person are extremely rare, they only occur in 0.09% of all uploads.

Conclusion

The quality checks have proven to be an effective and inexpensive means of filtering out obviously bad photos and avoiding subsequent rejection in manual quality control. With 7.5% of people having their photo automatically rejected at least once, we expect the quality checks to reduce the rejection rate in manual quality control by about a third (5 out of 15 percentage points).

Data used

The data comes from various Photo Collect instances from the spring of 2023. The analysis is based on about 23,000 photos.

‍