Can you match a photo to the discrete camera it was taken with, without metadata? [duplicate]

Question

Are there enough digital and/or analog difference in individual cameras/houses/lenses that photos will have fingerprint of sorts? For example discoloration of pixels, distortion from aberrations on lenses.

Imagine if you will that proving that an image (stripped of metadata) was taken with one specific camera would be critical to an investigation of some sort. If endlessly competent experts had full access to the camera and photo, and had every resource available, could they match a photo to a camera? Like a fired bullet to a gun.

When, or in what circumstance could a photograph be linked not only to a lens/house model or manufacturer but to one specific camera, distinguished from any other camera of the same model?

I guess a more general way of asking this question is: How much, and what information about the camera can a photograph capture?

Answer 1

Assuming the data has not been doctored in order to specifically spoof it's origin you could to a degree determine the camera/lens, depending on how much data you has available.

Cameras

If you had all the RAW data (in a format such as DNG) you could narrow it down quickly based on

• exact number of pixels
• how many pixels are masked at the edge of the frame

If you had a crop of the RAW data you could still look at

• bit depth (though some cameras let you choose)
• black point clipping (some manufacturers apply this to the RAW data)
• gaps in the histogram (Sony use a slightly nonlinear encoding that starts of counting 1,2,3 etc. then jumps to a coarser scale 104, 108, 112 etc), Canon's intermediate ISOs are implemented using digital arithmetic hence certain values will be absent.
• dynamic range / banding patterns (some sensors exhibit noticeable patterns in the shadows)
• Frequencies missing due to RAW "cooking" (e.g. noise reduction) I think Sony used to do this. This and other oddities can be revealed by looking at the Fourier transform.
• the properties of the CFA used (provided you had an image of a well known object / colour chart)

with this approach you're likly to be able to narrow it down to a particular sensor manufacturer (the camera itself leaves very little in the way of a fingerprint, so it would be hard to distinguish between a Sony sensor in a Pentax body and a Sony sensor in a Nikon body). As Canon make their own sensors you could identify a Canon based on characteristic banding patterns, or the number of masked pixels, but Canon use the same or very similar sensors in many different camera bodies.

With just a processed or resized image it would be much much harder. You could still attempt to measure the dynamic range (provided there was evidence of clipped highlights) though this could become unreliable due to the processing.

---

Lenses

Unlike sensors there are few dead giveaways (such as pixel count), however the lens used could be narrowed down by looking at by a combination of

• field of view (though this could be masked by cropping)
• out of focus highlights (number of aperture blades, Canon 50 f/1.8 is easy to spot with 5), spherical aberration (certain lenses are strongly over or undercorrected for SA which may be determined by looking for dark or light edges to oof highlights), catadioptric lenses are easy to spot due to donut bokeh.
• other quirks (swirly bokeh, astigmatism, coma)
• distinctive lens flare (see the 50mm f/1.8 images in this question)
• extreme parameters that are rare (ultra wide field of view / maximum aperture)

By combining information about the lens and sensor you can narrow things down further as some lenses can only be mounted on some bodies (without extensive modification).

Finally even with a resized, processed image you can sometimes identify a camera/lens based on attributes unique to the format. I was watching the independent film "Rubber" and within seconds I could tell it was shot with a Canon 5DmkII based on the depth of field at wide angle, due to the fact the sensor is larger than almost all available cine cameras (both film and digital) apart from 70mm monsters which would have been far outside the budget of this production.

Answer 2

In the strict sense, no, because the same digital numbers could have been produced in many ways, even completely artificially (not from measuring light). Therefore, unlike the grooves on a bullet or fingerprints, deliberate deception is easier to do and harder to detect.

If you can rule out deliberate deception, then I think it could be possible, depending on the characteristics of the scene and exposure.

First, just the image size can tell you a lot. Especially for high end cameras, the exact number of pixels in X and Y greatly narrows it down, possibly even to a particular model.

Other possibilities include looking for a noise signature in low-light regions of the image. Of course this is difficult if there aren't any or they are small. This is assuming you have the raw file (with metadata stripped). Compression artifacts will likely swamp small noise signatures.

Again assuming you have the raw image, just the Bayer pattern layout will eliminate many camera models. The number of scan lines per block, assuming TIFF-style format, also varies between models. So do other somewhat arbitrarily-chosen characteristics that don't matter to the photographer. Even the order of TIFF tags and the offset of the data fields into the file are identifiying characteristics. There are a lot of these little things that probably go a long way to narrowing down the model of camera.

There is a good chance that the sensor had some dirt somewhere. Just one obviously-off pixel could be a pretty conclusive if you can show that the particular camera in question has dirt at that location. Take a blank picture with the test camera, then look for any anomalies. Most likely you'll find a few. Especially if they can be traced to dirt on the sensor, that should be as good as a fingerprint. The chance of two cameras of the same model having even just two dirt spots in exactly the same place is astronomically small.

Answer 3

I found this question very fascinating and creative!!

I would think no, unless that particular sensor had some striking characteristic like a bad pixel or had dust on it that was not cleaned. Metadata can be faked and jpeg compression gets rid of a lot of stuff so I would say in terms of the legal system the most you could say was estimate the focal length of the lens, ISO and shutter speed (from knowing what is on the market).

I seriously doubt you could match it like a bullet can be matched.