Zero-Shot Learning Framework for 3D Texture–Geometry Defect Estimation in Mixed Fruits and Vegetables Using Signed Distance Fields

Objectives: The main goal of the research is to create a framework that is geometry motivated and can find and localize hidden three-dimensional defects with severity level supervision. Method: A zero-shot learning system called SDF-ZSL-DeTex is introduced that operates on dual Signed Distance Field representations. Canonical Geometry SDF is a model of defect-free morphology, whereas Defect Texture Deviation SDF is a model of local geometry-texture discrepancies to canonical structure. The Latent Attribute Discovery module automatically detects defect prototypes by using samples of severity-labeled samples. Zero-shot alignment loss is a structured separation imposed on healthy morphology and defect caused deviations. The framework gives both severity prediction and voxel-level localization predictions based on single implicit geometric reasoning. Findings: Experimental analysis of the Mango, Muskmelon, and Pomegranate datasets reveal the defect detection accuracy of 94.6%, the severity classification Macro-F1 up to 86.9%, and localization IoU up to 81.4%. Further testing on publicly available Potato and Tomato data sets attains a classification accuracy of 98.92% and 98.40% respectively. Findings are consistent with high levels of cross-produce generalization and localization performance as opposed to newer voxel-based and point-based three dimensional anomaly detection models. Novelty: This is demonstrated by the integration of canonical geometry modeling, deviation-based implicit representation, latent attribute discovery, and zero-shot alignment learning in a single dual Signed Distance Field architecture, to allow severity-dependent defect reasoning without semantic defect annotations. Representation based on continuous geometry facilitates powerful detection of unexplored patterns of defects in a variety of fruit and vegetable groups.