Scene Text Recognition (STR) focuses on extracting textual information embedded within natural scene images, playing a vital role in applications like document analysis, and automated translation systems. Despite its importance, STR remains a challenging task due to the diversity of text styles, orientations, and backgrounds in real-world images.

My work in STR aims to address several key challenges: improving the recognition of irregular and non-horizontal text arrangements, enhancing the handling of noisy, low-resolution images, and developing methods that generalize across languages. Current models often struggle with linguistic variations, diacritic characters, and diverse font styles, making them less applicable to non-English languages or real-world conditions.

To tackle these issues, my research emphasizes the creation of robust datasets and advanced models capable of handling multi-language text, diverse fonts, and environmental conditions. This includes exploring novel synthetic data generation techniques and innovative recognition architectures tailored for scene text's unique challenges.

Mitosis detection is crucial in the field of digital pathology, as it plays a vital role in diagnosing and prognosing cancer. This task involves identifying mitotic cells within histopathological images, which is challenging due to the high variability in cell morphology, overlapping cells, and the presence of cells that mimic mitotic features.

My research focuses on addressing these challenges by developing robust and generalizable models. Key objectives include reducing false positives, tackling domain shift caused by variations in scanners, staining protocols, and tissue types, and ensuring high performance on out-of-domain datasets.

Utilizing advanced techniques like domain adaptation, deep learning architectures, and comprehensive datasets, I aim to enhance the accuracy and reliability of mitosis detection systems. These improvements are essential for integrating computer-aided diagnostic tools into clinical workflows, ultimately aiding pathologists in achieving faster and more consistent diagnoses.

Nuclei instance segmentation is a critical task in the analysis of histopathology images, enabling the accurate identification and quantification of individual cell nuclei. This process is essential for understanding tissue morphology and aiding in disease diagnosis, particularly in cancer studies.

My MSc thesis focused on addressing the challenges in nuclei instance segmentation, including the detection of closely clustered nuclei and handling domain variations across different datasets. The study proposed novel methodologies like HR-YOLO, optimized for small object detection, and YOLO-U, a fusion of semantic segmentation and object detection techniques.

This research introduced standardized training and testing protocols, ensuring fair comparisons of segmentation models across widely-used datasets such as MoNuSeg, CoNSeP, and CPM17. These contributions aim to bridge the gap between algorithmic advancements and their practical applications in medical image analysis.

Semi-supervised learning (SSL) bridges the gap between supervised and unsupervised learning by leveraging both labeled and unlabeled data during training. This approach is particularly beneficial in scenarios where acquiring labeled data is expensive or time-consuming, while unlabeled data is abundant.

My research in SSL focuses on domain adaptation, tackling the challenge of training models that generalize effectively across diverse and unseen domains. This includes developing robust pseudo-labeling techniques and ensemble learning strategies to improve model reliability and performance in real-world applications.