Instance Segmentation

Unlike semantic segmentation, which labels all pixels of the same class identically, instance segmentation assigns a unique identity to each distinct object. Two overlapping or adjacent defects of the same class—such as scratches—receive separate masks with distinct instance IDs. This is critical for counting individual defects, measuring per-defect morphology, and triggering precise robotic rejection of single faulty units without discarding conforming products in the same field of view.

The core output is a binary mask for each detected instance, classifying every pixel as belonging to the object or the background at the original image resolution. Architectures like Mask R-CNN add a parallel segmentation head to a Faster R-CNN detector, generating masks via a Fully Convolutional Network. This enables precise area, perimeter, and circularity measurements for metrology applications, quantifying defect severity beyond simple bounding box dimensions.

Dominant architectures operate in two sequential stages:

• Stage 1: Region Proposal — A Region Proposal Network scans the image and proposes candidate bounding boxes likely to contain objects.
• Stage 2: Classification and Mask Generation — Each proposal is classified, its bounding box refined, and a pixel-level mask predicted in parallel. This decoupled approach achieves high accuracy but introduces latency, making it suitable for offline or near-real-time quality inspection rather than ultra-high-speed lines.

For real-time production line deployment, architectures like YOLACT and SOLOv2 perform instance segmentation in a single forward pass. YOLACT generates prototype masks and predicts per-instance linear combination coefficients, assembling final masks without a dedicated repooling step. These models trade marginal mask precision for inference speeds exceeding 30 frames per second, enabling inline defect segmentation on fast-moving conveyors.

Segmentation accuracy is evaluated using Intersection over Union computed directly on masks, often termed mask IoU. The standard COCO evaluation protocol measures Average Precision at multiple mask IoU thresholds from 0.50 to 0.95, providing a comprehensive view of segmentation fidelity. For manufacturing, a high IoU threshold ensures that defect boundaries are precisely delineated, preventing false scrap of borderline conforming material.

Instance segmentation inherently resolves partial occlusion by predicting separate masks for overlapping objects. The model learns to infer the full extent of a partially hidden defect based on visible contours and contextual cues. This capability is essential for inspecting densely packed components—such as semiconductor wafers or assembled circuit boards—where defects frequently overlap or touch, confounding simple bounding box detectors.

A pixel-level classification task that assigns a class label to every pixel in an image. Unlike instance segmentation, it does not distinguish between individual objects of the same class. For example, all pixels belonging to overlapping scratches would receive the same label, merging them into a single amorphous region. This is useful for quantifying total defect area but insufficient for counting distinct defects.

A task that localizes objects by predicting bounding boxes and class labels. While it distinguishes individual instances, it provides only a coarse rectangular enclosure. For irregularly shaped manufacturing defects like cracks or delamination, a bounding box includes significant non-defect background pixels, making precise dimensional measurement impossible. Instance segmentation extends this by adding a pixel-precise mask.

A unified task that combines semantic and instance segmentation. It assigns every pixel a class label and, for countable objects, a unique instance ID. Background classes like 'conveyor belt' or 'metal surface' are treated semantically, while foreground classes like 'scratch' or 'dent' receive instance-level masks. This provides a complete, holistic scene understanding for total quality analysis.

The standard metric for evaluating instance segmentation models. It measures the area under the precision-recall curve, averaged across all classes and Intersection over Union (IoU) thresholds. A prediction is considered a true positive only if the predicted mask achieves sufficient overlap with the ground truth mask. mAP@50:95 is the rigorous COCO challenge standard, averaging performance across IoU thresholds from 0.50 to 0.95.

A metric computing the ratio of the area of overlap to the area of union between a predicted mask and the ground truth mask. An IoU of 1.0 indicates perfect overlap. In instance segmentation, an IoU threshold—typically 0.50 or 0.75—determines whether a prediction is a true positive. This pixel-level metric is far more stringent than bounding box IoU, penalizing shape inaccuracies.

A seminal deep learning architecture that extended Faster R-CNN by adding a parallel branch for predicting segmentation masks on each Region of Interest. It introduced the concept of decoupling classification and mask prediction, where mask generation is class-agnostic. This two-stage detector remains a robust baseline for industrial defect instance segmentation, offering high accuracy at the cost of inference speed.