I want to post my testing results in case somebody also interested in the topic.
I inspected the apps from the review kindly suggested by Arthur and built a following ranking:
ScanSpeeder. Showed the best results, it recognized all the photos from the non-standard canvas with distorted edges. In fact the canvas was the dirty table with a cover of grey-brown color, which is extremely difficult for image-recognition software. It recognized it all, with 90% precision. Excellent.
AutoSplitter. Showed worse results than ScanSpeeder, from 5 images on the canvas it recognized 3 and partly the fourth. Not bad.
In fact there is nothing more I can add to ranking.
ImageSplit from CaptionPro did not detected nothing, besides it is heavily advertised like professional software. It is useless.
VueScan is useless either, it does not allow processing arbitrary images, only captured from scanner.
Photoshop I didn't tested, I trust it can do such simple tasks, he-he, but it lackas automation.
RansenScan didn't detected nothing either, or I didn't understand the principles of its functioning.
Regarding GIMP and its DivideScannedImages script I wasn't successfull in installing it, it is 7 years old so it may be even not applicable to current GIMP at all. I will try and if succeeded, will report here.
I want to stress here, that I was interested in an automated way of doing splitting, so I didn't check manual way, only recognition algorithms. Each of the above apps has manual selection mode of extracting images from canvas.