The testcases, after rejudgement, are harder than before. So I really wonder if simply randomization could get AC.

I first wrote a program WHOLLY random, and got WA at 4.
Then I added the checking part of my program. When one-round randomization failed, I will start another, and so on. Then I got TLE at 14, 15 or 19 (I have tried about 10 times).
Finally, I realised that I must modify the random part. So I made a improvement: when I try to find the value of one pixel, I will take the PROCESSED pixel which could have the impression on that pixel into consideration. And I will adjust the random argument basing on these infos. This time I got AC.

Wish my method will be helpful.