Abstract:Semi-supervised domain adaptation (SSDA) has been extensively researched due to its ability to improve classification performance and generalization ability of models by using a small amount of labeled data on the target domain. However, existing methods cannot effectively adapt to the target domain due to difficulty in fully learning rich and complex target semantic information and relationships. In this paper, we propose a novel SSDA learning framework called semantic regularization learning (SERL), which captures the target semantic information from multiple perspectives of regularization learning to achieve adaptive fine-tuning of the source pre-trained model on the target domain. SERL includes three robust semantic regularization techniques. Firstly, semantic probability contrastive regularization (SPCR) helps the model learn more discriminative feature representations from a probabilistic perspective, using semantic information on the target domain to understand the similarities and differences between samples. Additionally, adaptive weights in SPCR can help the model learn the semantic distribution correctly through the probabilities of different samples. To further comprehensively understand the target semantic distribution, we introduce hard-sample mixup regularization (HMR), which uses easy samples as guidance to mine the latent target knowledge contained in hard samples, thereby learning more complete and complex target semantic knowledge. Finally, target prediction regularization (TPR) regularizes the target predictions of the model by maximizing the correlation between the current prediction and the past learned objective, thereby mitigating the misleading of semantic information caused by erroneous pseudo-labels. Extensive experiments on three benchmark datasets demonstrate that our SERL method achieves state-of-the-art performance.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve a key problem in semi - supervised domain adaptation (SSDA): **How to effectively utilize a small amount of labeled data and a large amount of unlabeled data in the target domain to fully learn complex target semantic information and achieve better domain alignment**. Specifically, existing SSDA methods face challenges in the following aspects: 1. **Insufficient learning of target semantic information**: - Existing methods are difficult to comprehensively capture the rich semantic information in the target domain and the relationships between samples, resulting in poor performance of the model on the target domain. 2. **Feature representation biased towards the source domain**: - Since there is a large amount of labeled data in the source domain, the feature representation of the model is easily biased towards the source domain, thus affecting its generalization ability on the target domain. 3. **Impact of pseudo - label noise**: - Pseudo - labels in the target domain may be noisy, which will cause the model to learn wrong semantic information and then affect the model performance. To solve these problems, the author proposes a new framework - **Semantic Regularization Learning (SERL)**. SERL enhances the model's learning of target - domain semantic information through three regularization techniques: - **Semantic Probability Contrastive Regularization (SPCR)**: - Helps the model learn more discriminative feature representations from a probabilistic perspective and reduces the influence of low - confidence samples through adaptive weights. - **Hard - sample Mixup Regularization (HMR)**: - Uses simple samples as guidance to mine the potential knowledge in hard samples and helps the model learn more complex semantic distributions. - **Target Prediction Regularization (TPR)**: - Minimizes the influence of noisy pseudo - labels to ensure that the model learns correct semantic information. These techniques work together to enable the model to better adapt to the actual distribution of the target domain and significantly improve its classification performance and generalization ability. ### Summary The main contribution of this paper is to propose a new SSDA framework SERL, which makes full use of the semantic information of the target domain through semantic regularization strategies, solves the deficiencies of existing methods in target - domain semantic learning, and thus achieves more effective cross - domain adaptation.

Source-free Semantic Regularization Learning for Semi-supervised Domain Adaptation

When Masked Image Modeling Meets Source-free Unsupervised Domain Adaptation: Dual-Level Masked Network for Semantic Segmentation

Semi-supervised Domain Adaptation for Semantic Segmentation

Learning from Different Samples: A Source-free Framework for Semi-supervised Domain Adaptation

Local Weight Coupled Network: Multi-Modal Unequal Semi-Supervised Domain Adaptation.

SSF-DAN: Separated Semantic Feature Based Domain Adaptation Network for Semantic Segmentation

Perception and Semantic Aware Regularization for Sequential Confidence Calibration

Semi-Supervised Domain Adaptation via Adaptive and Progressive Feature Alignment

CSCL: Critical Semantic-Consistent Learning for Unsupervised Domain Adaptation

Semi-supervised Domain Adaptation with Subspace Learning for Visual Recognition

Alleviating Semantic-level Shift: A Semi-supervised Domain Adaptation Method for Semantic Segmentation

The Last Mile to Supervised Performance: Semi-Supervised Domain Adaptation for Semantic Segmentation

Consistency Regularization for Generalizable Source-free Domain Adaptation

SS-ADA: A Semi-Supervised Active Domain Adaptation Framework for Semantic Segmentation

Semantic Concentration for Domain Adaptation

Pixel-Level Domain Adaptation: A New Perspective for Enhancing Weakly Supervised Semantic Segmentation

Unsupervised Domain Adaptation for Remote Sensing Image Semantic Segmentation Using Region and Category Adaptive Domain Discriminator

Regularizing Proxies with Multi-Adversarial Training for Unsupervised Domain-Adaptive Semantic Segmentation.

Regularizing Deep Networks with Semantic Data Augmentation

Self-Supervised Representation Learning with Meta Comprehensive Regularization

Learning Invariant Representation with Consistency and Diversity for Semi-supervised Source Hypothesis Transfer.