Research On Decoupling-Training-Based Semi-Supervised Image Semantic Segmentation Algorithm

Semantic segmentation task is an important direction in the field of computer vision.However,the huge cost of obtaining pixel-level semantic segmentation labels hinders the development of semantic segmentation task.Semi-supervised semantic segmentation algorithm can reduce the labeling cost of semantic segmentation tasks by using a small amount of labeled data and a large number of unlabeled images to train the semantic segmentation model.The existing semi-supervised semantic segmentation methods are usually based on the "teacher-student" model paradigm,combined with the consistent learning theory,and use the pseudo-label produced by the teacher model to supervise student model with strong interference.However,the class-imbalance problem makes the pseudo-label generated in the semi-supervised training process more inclined to the head category.The unbalanced pseudo-labels are used to supervise the model training further aggravates the classimbalance problem of the model and affects the model performance.In order to solve this problem,this paper proposes a decoupling semi-supervised semantic segmentation algorithm inspired by decoupling training,which cuts off the gradient connection between the encoder,decoder and the semantic segmentation head,so that the classimbalance problem cannot propagate in the semantic segmentation model.The main research work is as follows:1、For the encoder,this paper uses all labeled data and unlabeled images to train an encoder with strong robustness based on semi-supervised method.2、For the decoder,this paper uses pixel-level re-weighted loss training to balance the decoder.In addition,in order to make full use of unlabeled images to improve the performance of the decoder,this paper proposes to use the future student model to generate pseudo-label for the decoder of the teacher model.3 、 For the semantic segmentation head,this paper uses the shared semantic segmentation head to connect the teacher model and the student model.This nongradient connection method not only transfers the balance information of the teacher model to the student model,but also restricts the consistency of the feature distribution between the teacher model and the student model,which is beneficial to the training of the student model.4 、 Finally,this paper proposes a multi-level entropy equalization sampling algorithm,taking into account the balance of the number of samples inter-class and the entropy balance of samples intra-class.The pixel-level feature obtained from sampling are used to update the shared semantic segmentation head,so that it has a more balanced classification ability.In order to verify the effectiveness of the algorithm proposed in this paper,extensive experiments have been carried out on Pascal VOC 2012 dataset and Cityscape dataset.The ablation experiment on the method proposed in this paper shows the effectiveness of the proposed method.Comparing the proposed method with the existing semi-supervised semantic segmentation method,the proposed method has achieved the best experimental results in the above two data sets.