Research On Semi-supervised Object Segmentation Based On Pseudo-label Training

Semantic segmentation is one of the most basic tasks in computer vision,which requires detecting of pixel semantic categories of each object in an image.Since its performance depends on the amount of available annotation data,while labeling pixellevel precise information incurs huge annotation time and labor costs.In order to alleviate this problem,the semi-supervised object segmentation method based on pseudo-label training has attracted wide attention in recent years due to its ability to use a large amount of unlabeled data to improve the accuracy of the model.When applying the semantic segmentation method based on semi-supervised learning to a specific task,a pre-trained model needs to be obtained on a similar benchmark dataset first.Subsequently,the task segmentation model is trained on the taskannotated data.The predicted results on unlabeled data obtained using the task segmentation model are used as pseudo-labels to obtain additional training data.However,when this method is applied to a novel domain,such as leakage detection,the semantic information in the task data differs significantly from that in the benchmark dataset.Directly using a specific benchmark dataset for pre-training and then training the task segmentation model on the task-annotated data will result in inaccurate predictions on unlabeled data.The pseudo-labels will contain a large amount of misclassification noise.Suppose these noisy pseudo-labels are directly used to train the task segmentation model without screening.In that case,the model will learn incorrect information,and as the training iterations increase,the model will accumulate errors caused by incorrect information,eventually leading to overfitting.In view of the above problems,the paper carried out the following research:(1)To solve the problem of pseudo-label noise caused by semantic information difference,the paper proposes a semi-supervised object segmentation algorithm via reliability-aware.The core of the algorithm is to quantify the reliability of pseudo-labels.The reliability of the pseudo-labels was comprehensively evaluated by the confidence of the model and the consistency between the models,and the pseudo-labels were filtered according to the reliability score to improve the accuracy of the model by reducing the amount of noise in the training process.(2)Aiming at the problems of error accumulation and overfitting under the influence of wrong classification,the paper proposes a semi-supervised object segmentation method based on reliability-aware combined with two-stage fine-tuning.Instead of mixing pseudo-labels and real labeled samples as training data,pseudo-labels and labeled samples are fine-tuned separately.By giving the model unbiased input,the staged finetuning strategy can simply and effectively stimulate the network’s correction ability,making the network recognize the existence of errors and correct the accumulation of errors to avoid model overfitting and improve the robustness.At the same time,the implementation of the fine-tuning strategy does not need additional calculations and data.(3)The paper researches a semi-supervised object segmentation algorithm based on pseudo-label training in a novel domain,a novel domain dataset with data semantics different from the benchmark dataset is needed to verify the validity of the proposed method.For this purpose,a novel domain dataset named NDOS,which contains 330 detection images for industrial detection,is provided and made public.It differs from the existing benchmark dataset in semantics and can provide a new benchmark for testing novel domain semantic segmentation methods.(4)The reliability-aware and two-stage fine-tuning methods proposed in the paper are experimentally verified on the novel domain dataset.The experimental results show that the proposed methods can effectively improve the model’s accuracy.