BiCG-Net scores the target, not the frame: differencing 11 teacher readings within one image lifts zero-shot transfer PLCC to 0.8416
Synopsis
When a target occupies only a small part of the frame, whole-image quality scores are dominated by background, and public benchmarks carry no target-level human opinion scores. BiCG-Net expands the target box into a nested patch sequence, reads it with 11 frozen image quality models, and subtracts two readings from the same image to cancel the target term and teacher bias by construction. On CSIQ and LIVEC the student reaches PLCCs of 0.9009 and 0.9133 against human MOS, above both the best single teacher and the teacher consensus; frozen and transferred zero-shot from CSIQ to LIVEC it reaches 0.8416, 0.1947 above consensus.
Interpretation
Image quality assessment moves from the whole frame to a designated target region, scoring the object inside the target box rather than the entire image. Mainstream no-reference quality assessment treats the whole frame as one homogeneous quality unit, so when the target covers a small fraction of the frame the aggregated score is dominated by background, track ballast and vegetation. On the self-built railway dataset the best single teacher reaches a PLCC of 0.8230 and the teacher consensus 0.8642, while BiCG-Net reaches 0.9855; that dataset is scored against a weak label, not human MOS.
The method expands the target box outward into a nested patch sequence, has each teacher read every patch, and infers target quality from how the readings change as the patch grows. Each teacher reading is written as a sum of five terms: target quality, teacher bias, area effect, background content and noise, none of them directly observed. Subtracting two readings from the same image cancels the target term and teacher bias because both are invariant within an image, leaving the area effect, background difference and noise; a geometric encoder estimates and removes the area effect.
The student network does not imitate any single teacher; it learns a mapping from the relative structure of teacher readings to the quality label. Conventional multi-teacher aggregation takes an arithmetic mean of teacher outputs, which carries teacher bias and background terms into the supervision signal. On CSIQ and LIVEC the student reaches PLCCs of 0.9009 and 0.9133 against human MOS, above the best single teacher and the teacher consensus; frozen and transferred from CSIQ to LIVEC it reaches 0.8416 against a consensus of 0.6469.
Perspective
The demonstrated scope covers two public benchmarks with human MOS, CSIQ and LIVEC, and a self-built railway dataset of 2,816 images; on the public benchmarks a centered box with a 1/9 area ratio serves as a pseudo-target, so the target-level task degenerates to whole-image assessment. Results on the self-built dataset are measured against a weak label generated from teacher consensus, not human MOS. The authors propose adding a subset of human MOS annotations to the self-built dataset to measure how the weak label correlates with human judgment.
The 0.9855 on the self-built dataset reflects fidelity to weak supervision, not direct agreement with human perception. Cross-dataset transfer is asymmetric; the authors hypothesize this reflects the source manifold's distortion coverage, but the present experiments do not separate it from dataset size or MOS scale. The dispersion gate changes PLCC by 0.0000 on all three datasets, and the authors position it as a diagnostic pathway rather than an accuracy module.