Cell detection method

Abstract

The invention discloses a cell detection method which comprises the following steps: s1, manually labeling point labeling information on a test image, and extracting a plurality of labels from the point labeling information; s2, performing joint learning on the plurality of labels to obtain a plurality of teacher models; s3, performing learning optimization on the weights of the plurality of teacher models; s4, based on the weight of the teacher model obtained in the step s3, generating a weighted label, and using the weighted label to train a single student model; and s5, completing cell detection test based on the trained student model obtained in the step s4 to obtain a detection result. Based on a knowledge distillation model method, the cell detection performance is improved by more efficiently utilizing point labeling information.

Description

technical field [0001] The invention relates to the application field of computer neural network technology, in particular to a cell detection method. Background technique [0002] Image semantic segmentation technology based on deep convolutional neural network is widely used in cell detection problems. Compared with traditional vision algorithms, image semantic segmentation technology based on deep convolutional neural network has achieved significant performance improvement. [0003] Some classic deep convolutional neural network structures, such as FCN (fully convolutional neural networks), U-net and Segnet, have become benchmark solutions in cell detection problems. However, due to the intricate scenes in the cell detection problem, such as cohesive cells, background noise, and variable cell size and shape, the performance of these image semantic segmentation technologies based on deep convolutional neural networks still needs to be further improved on cell detection p...

AI Technical Summary

Problems solved by technology

However, due to the intricate scenes in the cell detection problem, such as cohesive cells, background noise, and variable cell size and shape, the performance of these image semantic segmentation technologies based on deep convolutional neural networks still needs to be further improved.

[0004] Some current cell detection techniques cannot further effectively improve the performance of cell detection because they cannot make full use of point annotation supervision information.

Images