I would really like to see if there would be a way to work on a hash algorithm that was not focused on nonreversability but more focused on uniformity to avoid the (small change) -> (same hash) problem.

That wouldn't change anything. All a hash does is mix the bits of source data to attempt to come up with a "unique" number that identifies that data, but in doing so, it throws away information. In throwing away information, you insure that your hash isn't really unique. It is just hard to make a file that has the same hash. Hard, but not impossible. If you have a 4byte perfect hash, and hash a 5 byte string, there are 256 4byte prefixes depending on the value of the 5th byte. Without knowing what the 5th byte's value is, you can't figure out the value of the first 4 bytes by using the hash function. Now, compound that problem when you add more bytes.

To give an example from school, we were shown how you can take a c++ "hello world" program and make it say "goodbye world" but have the same CRC32 hash value. Simple, replace the string, and toss a /* garbage comment */ at the end of the file. Fill the comment with with bits till it comes up with the same hash. Without a visual inspection by a human, you'd never know the file'd been tampered with. You can compile it to verify that it works, and the same CRC32 hides the fact that you did tamper with it. You can do the same thing for most files. You can edit a JPG image, and then fill the comments with garbage till the hash is the same. With a bitmap you could make some big edits, then twiddle around with the least significant bits of every pixel, as the human eye isn't going notice the differences. Now you have an image and a shopped image that both have the same SHA-256 hash. A good shop is hard to tell from the origional, and they have the same hash, so what image is the one you were looking for?