New training with 588 images 17500 steps against 453 and 16000 on version 2. Which I merged with another model at 350 images and 8000 steps for optimal results. This version required more than 100 model combination tests, all trained by me of course.