Objective: Due to varying subjective discrimination scales and excessively long network computational responses, During the plasma preparation process, there is a tendency for suspected hemolytic plasma to be misjudged and non suspected hemolytic plasma to be medically discarded, It poses great risks to the patient's life safety. Methods: Regarding the issue of minimal differences in artificial vision between suspected hemolysis and non suspected hemolysis in biological samples of human blood plasma at critical states,This article proposes a bounded method based on deep learning and color changing ink concepts. Results: Using an improved U-Net network for image segmentation, Introducing optimized CA module, batch normalization, and filling module, to solve the problems of estimation mean shift, low computational efficiency, and insufficient receptive field of view in spatial mapping relationships. Conclusion: Using the color changing ink limit method for classification discrimination, While ensuring the accuracy of plasma sample recognition, Improved the computational efficiency of plasma discrimination and reduced discrimination time. The model was validated and compared using a self collected sample dataset. The results indicate that the plasma discrimination accuracy of our model is superior to other discrimination models.