Yiming Li

Yiming Li

Auto-generated profile, based on publicly available information

Yiming Li

National Key Laboratory of Radar Signal Processing, Xidian University, Xian, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
12 September 2024 in IEEE Transactions on Geoscience and Remote Sensing
Verified
Journal Article
Closed access
YSYu Shi
Yu Shi
LDLan Du
Lan Du
YGYuchen Guo
Yuchen Guo
YDYuang Du
Yuang Du
YLYiming Li
Yiming Li
This paper delves into the topic of unsupervised domain adaptation (UDA) by transferring knowledge from rich labeled optical domain to unlabeled SAR domain, tackling the issues faced by deep-learning-based SAR ship classification methods that rely on abundant labeled SAR images. Typical UDA methods usually extract domain invariant representations (DIR) between two domains. However, due to the prominent differences in imaging mechanisms between optical and SAR images, the discriminative characteristics of same classes across domains may vary. Feature representation guided by labeled optical images therefore suffers from a particularly serious source-bias problem, making DIR difficult to be extracted. Moreover, capturing the category structure of the target domain is crucial for classification tasks. To solve the above challenges, this paper proposes a UDA framework for SAR ship classification via progressive feature alignment between optical and unlabeled SAR domains, gradually aligning two domains across domain-, and class-levels. At the domain-level, to reduce the transfer difficulty stemming from the prominent differences between SAR and optical images, feature calibrated domain alignment (FCDA) is presented to achieve accurate DIR extraction. FCDA combines the reconstruction and the consistency constraints of different perturbed versions of the same image to calibrate the optical-bias representation into the features of unbiased toward specific domain. At the class-level, we proposed feature enhanced class alignment (FECA) to capture the fine-grained category structure of the SAR domain. FECA incorporates pseudo-labels-based cross-domain contrastive learning for intra-class compactness as well as inter-class separation among cross-domain categories, along with a consistency learning approach to enhance the class structure of SAR domain. The experimental results indicate that our method achieves exceptional performance in unsupervised classification of SAR ships.

13 March 2025 in IEEE Transactions on Aerospace and Electronic Systems
Verified
Journal Article
Closed access
YLYiming Li
Yiming Li
LDLan Du
Lan Du
HLHuayue Liu
Huayue Liu
YGYuchen Guo
Yuchen Guo
Class-incremental learning based on convolutional neural networks for synthetic aperture radar (SAR) targets has garnered wide attention recently. Unlike incremental SAR targets classification, incremental SAR targets detection and classification requires the model to detect and classify old targets while learning current ones, which is challenging due to the absence of old targets in the incremental large-scale data. To address the above issues, this paper proposes a class-incremental SAR ship detection and classification method, which combines exemplar replay (ER) and knowledge distillation (KD) to facilitate the learning of the incremental model. Specifically, we propose a context-robust ER method to retain precise scene context between replayed old targets and clutter in the incremental data. By fully leveraging the scene characteristics of the ship targets, context-robust ER method combines a local context-aware strategy with sea-land segmentation to pinpoint regions with the highest likelihood of being sea areas. Then, it replays exemplars within these regions, effectively reducing context-bias issues. In addition, a multi-granularity KD method is introduced, which transfers the knowledge learned by the old detector to the current detector progressively. Multi-granularity KD method combines an object mask strategy with spatial-channel attention mechanisms to constrain the current detector to focus on the most important information from the old detector. Experiments conducted on SRSDD-v1.0 dataset indicate that our method achieves satisfactory performance in incremental detection and classification of SAR ships.

18 June 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Verified
Journal Article
Open access
YLYiming Li
Yiming Li
LDLan Du
Lan Du
JCJian Chen
Jian Chen
JSJialun Song
Jialun Song
ZWZilin Wang
Zilin Wang
YGYuchen Guo
Yuchen Guo
Current synthetic aperture radar (SAR) automatic target recognition (ATR) methods primarily focus on closed set recognition, i.e., they typically assume that the training set covers all target classes that appear in the test set. However, this assumption may not hold in practical scenarios, where it's inevitable to encounter high-value targets whose classes are absent from the training set. In this article, we propose an open set recognition (OSR) based on the encoding-conditional decoding network with reject threshold adaptation (ECTA) for SAR targets. The ECTA consists of four parts: 1) the encoding; 2) classification; 3) conditional decoding; and 4) reject threshold calculation modules. The encoding module learns the latent representation of the training samples, which is then fed to the classification part to predict the class labels. The conditional decoding module is trained to produce matching reconstructed samples by combining the latent representation with the label corresponding to the target class while producing nonmatching reconstructed samples by combining the latent representation with the label corresponding to a random other target class. Through the proposed reconstruction strategy, the conditional decoding module gives an approximate description of the known and unknown class reconstruction error distributions, building the foundation for the reject threshold calculation module. The reject threshold calculation module fits the extreme value distributions of matching and nonmatching reconstruction errors based on extreme value theory (EVT) and finally obtains satisfactory reject thresholds adaptively by minimizing the probability of false positives and false negatives. Experiments conducted on two measured datasets indicate that ECTA outperforms state-of-the-art OSR methods, such as DIAS (+0.01% $\sim$ 7.36% and +2.1% in F1-score on MSTAR against varying openness and FUSAR-Ship respectively).

20 October 2025 in IEEE Geoscience and Remote Sensing Letters
Verified
Journal Article
Closed access
YLYiming Li
Yiming Li
YGYuchen Guo
Yuchen Guo
LDLan Du
Lan Du
HMHongyang Ma
Hongyang Ma
Existing SAR ship detection and classification methods based on deep learning have achieved notable performance under the close-set assumptions. However, SAR object detection and classification tasks inherently operate in open-set scenarios, which naturally involve the potential presence of unknown non-cooperative targets. Closed-set methods cannot identify these unknown targets, often misclassifying them as known classes or background, leading to significant errors. Therefore, this paper proposes an unknown-target mining-based open-set SAR ship detection and classification method UTM-YOLOX, which preserves the performance for known targets while enhancing the ability to detect and identify unknown targets. Firstly, class-agnostic localization quality estimation (CLQE) is utilized in the regression heads of YOLOX to improve the detection of unknown targets. Unlike binary foreground-background classification, CLQE focuses on learning the quality of target localization, providing richer information for target detection. Secondly, an open-set classifier is proposed for classification heads of YOLOX to identify the unknown targets and known classes, which includes unknown class discriminant learning (UCDL) and contrastive clustering learning (C2L). UCDL utilizes the detected proposals of suspected unknown targets into the training of the open-set classifier, enabling the model to effectively identify unknown targets, while C2L is utilized to improve the discriminative ability between known and unknown targets. Experiments on SRSDD-v1.0 demonstrate that UTM-YOLOX achieves satisfactory performance. The code is available at https://github.com/LiYiming98/UTM-YOLOX.