A merge-split algorithm based on study property of neural net and sum aggregate is presented. Its merit is that its merge is disconnected a little and freedom satisfies the adaptability, and it is realized during learning rather than stipulating in advance. Furthermore, the method of fitting edge of uncontinuous image with neural net is given. The method can carry on the fit at self-learning, using its learning function fully under the circumstances without the mathematical function form of edge. Some shortages of former marginal fit methods can be overcomed. 