CSBC: Class and Scale Balanced Semi-Supervised Blood Cell Detection

Blood cell detection plays a vital role indiag nosing hematological diseases with microscopic images. However, conventional fully-supervised methods require large-scale annotated images, which are costly and labor intensive to obtain. While Semi-Supervised Object Detection (SSOD) methods can reduce the annotation cost by leveraging unlabeled images, most existing methods de pend on specific detector architectures, and their performance remains limited in blood cell images due to distinct class-level and scale-level imbalance. To tackle these challenges, we propose a Class and Scale Balanced semi supervised framework for Blood Cell detection (CSBC), which is compatible with various CNN-based detectors. It consists of three key components: First, a Dual-view High-quality Instance Selection (DHIS) strategy constructs a dynamic class-wise instance pool by selecting reliable pseudo-labels based on classification confidence and localization consistency between two detectors. Second, a Class and Scale Balanced Resampling (CSBR) module incorporates a novel balancing factor combining class frequency and object-scale information to adaptively adjust class-level sampling probability, thereby enhancing the representation of minority categories and small objects. Third, a Background-Aware Pasting Strategy (BAPS) is employed to preserve image structure by pasting sampled instances into background regions. Extensive experiments on the BCCD and PBC datasets demonstrate that CSBC consistently outperforms ten state-of-the-art SSOD methods under both Faster R-CNN and RetinaNet detectors, confirming its effectiveness and generalizability in blood cell detection.