Three-Level Fish Iris Image Annotatability Assessment for Pre-Annotation Quality Control Based on Boundary Traceability

Fish iris images provide morphological information for biometric recognition and biological observation, but underwater imaging often introduces blur, reflection, occlusion, low contrast, and boundary ambiguity. Conventional no-reference image quality assessment provides global perceptual scores but cannot determine whether local iris boundaries are reliably traceable for annotation. We formulate fish iris quality control as a three-level annotatability assessment task for pre-annotation screening, enabling direct annotation, further review, or exclusion/reacquisition. Under limited supervision, the framework learns complementary quality evidence from 53-dimensional handcrafted quality descriptors and 384-dimensional embeddings extracted from a frozen pretrained MANIQA (Multi-Dimension Attention Network for No-Reference Image Quality Assessment) model. LateAvg performs lightweight decision-level integration by equally averaging branch posterior probabilities. On 1500 self-collected underwater images evaluated by fish-group five-fold cross-validation, the framework achieves 0.8574 accuracy and 0.8510 Macro-F1. It achieves the highest mean Accuracy, Macro-F1, and Weighted-F1 among evaluated methods, outperforming scalar IQA baselines such as NIQE-score and TOPIQ-score while remaining statistically comparable to trainable integration and late fine-tuning approaches. These results support boundary-traceability-based pre-annotation screening. With cached quality evidence, the framework introduces only a lightweight decision stage to support efficient annotation-resource allocation in biological dataset construction.