Diffusion Planner Model for Surface Crack Detection of Quick-Frozen Dumplings in Low Temperature Production Lines

Source abstract

Surface cracks on quick-frozen dumplings are difficult to detect on low-temperature production lines because thin and discontinuous evidence is frequently obscured by frost, flour particles, packaging-film reflections, and nonuniform tunnel illumination. This study proposes a Diffusion Planner that treats inspection as constrained latent crack-path planning rather than direct texture classification or full-image generation. A temperature-normalized encoder first reduces appearance drift across freezer conditions, after which a contour-conditioned diffusion planner connects locally interrupted crack evidence and a compact decoder produces the final pixel-level mask. The planner is restricted to sparse task-relevant trajectories, while a confidence gate suppresses paths that conflict with normal pleat geometry. A production-line dataset containing 18,640 images and 42,310 annotated crack instances across four freezing-temperature bands was used for evaluation with product-sequence and acquisition-session separation. The proposed method achieved 95.6% precision, 94.2% recall, and a 94.9% F1 score, while reducing mean boundary error by 41.3% relative to the strongest conventional detector. At 1024 by 768 resolution, average inference latency was 24.8 ms on an industrial GPU, supporting a 1.2 m/s conveyor. Under heavy frost, diffusion planning increased recall by 3.7 percentage points, and temperature normalization reduced false alarms on legal folds by 28.6%. These results show that process-conditioned latent trajectory planning can provide accurate, interpretable, and deployment-oriented crack inspection for frozen food manufacturing.