You Only Look Once (YOLO)

Unlike two-stage detectors that first propose regions and then classify them, YOLO applies a single neural network to the full image. The architecture divides the input into an S × S grid; each grid cell predicts B bounding boxes, confidence scores for those boxes, and C class probabilities simultaneously. This end-to-end design eliminates the computational bottleneck of a separate region proposal network, enabling inference at 45+ frames per second on standard hardware.

YOLO imposes a strong spatial constraint by assigning each object's detection responsibility to the grid cell containing its center point. Each cell predicts a fixed number of bounding boxes, but only the box with the highest Intersection over Union (IoU) with the ground truth is selected as responsible. This spatial encoding forces the network to learn contextual features from the entire image rather than isolated regions, dramatically reducing background false positives compared to sliding-window approaches.

Modern YOLO variants (v3+) incorporate Feature Pyramid Networks (FPN) to detect objects at multiple scales. The architecture constructs a feature pyramid by:

• Extracting feature maps at three different resolutions from the backbone
• Applying 1×1 convolutional layers to reduce channel dimensions
• Upsampling higher-level semantic features and concatenating them with lower-level fine-grained features This enables simultaneous detection of both large structural defects and microscopic surface anomalies on the same production line.

YOLO uses pre-defined anchor boxes—bounding box priors with specific aspect ratios and scales—to improve localization accuracy. Instead of predicting absolute coordinates, the network predicts offsets relative to these anchors. The anchor dimensions are determined by running k-means clustering on the training dataset's ground truth boxes, using IoU distance rather than Euclidean distance. For manufacturing, this means anchors can be optimized for common defect shapes like scratches (high aspect ratio) or dents (square aspect ratio).

YOLO's custom backbone, Darknet, is designed for efficient feature extraction with 53 convolutional layers organized into residual blocks. Later versions integrate Cross Stage Partial Networks (CSPNet) to reduce computation by partitioning feature maps and only processing a subset through dense blocks. Key characteristics:

• Batch normalization after every convolutional layer for training stability
• Leaky ReLU activations to prevent dead neurons
• Global average pooling replacing fully connected layers to reduce overfitting This backbone achieves a favorable balance between FLOPs and accuracy for edge deployment.

YOLO's training objective is a compound loss function with three components:

• Localization loss: Sum-squared error on bounding box coordinates (x, y, w, h), scaled by λ_coord=5 to prioritize accurate positioning
• Confidence loss: Binary cross-entropy for objectness score, with λ_noobj=0.5 to down-weight loss from empty grid cells
• Classification loss: Categorical cross-entropy for class probabilities, applied only when an object is present This balanced loss prevents the gradient from being dominated by the overwhelming number of grid cells containing no objects, a critical consideration for defect detection where defective samples are rare.

The foundational deep learning architecture that powers YOLO's feature extraction backbone. CNNs apply convolutional filters to automatically learn spatial hierarchies—from simple edges in early layers to complex defect textures in deeper layers.

• Backbone networks: YOLO typically uses CSPDarknet or EfficientNet as its CNN backbone
• Feature pyramids: Multi-scale feature maps enable detection of defects at various sizes
• Weight sharing: Convolutional kernels scan the entire image, making CNNs translation-invariant

A critical post-processing step that eliminates redundant bounding box predictions. When YOLO predicts multiple overlapping boxes for the same defect, NMS selects the one with the highest confidence score and suppresses others exceeding an IoU threshold.

• IoU threshold: Typically set between 0.5 and 0.7 for manufacturing applications
• Soft-NMS variant: Decays scores rather than zeroing them, improving recall for clustered defects
• Per-class NMS: Applied independently for each defect category to prevent cross-class suppression

The primary localization metric measuring overlap between predicted and ground truth bounding boxes. Calculated as the area of intersection divided by area of union, IoU directly determines whether a detection counts as a true positive.

• IoU ≥ 0.5: Standard threshold for a correct detection in manufacturing quality inspection
• IoU = 1.0: Perfect prediction, extremely rare in practice
• GIoU and CIoU: Advanced variants that account for box shape and center-point distance, often used as YOLO's regression loss function

The definitive evaluation metric for YOLO models, summarizing detection performance across all defect classes and IoU thresholds. [email protected] measures precision-recall area at IoU ≥ 0.5, while [email protected]:0.95 averages across ten IoU levels from 0.5 to 0.95.

• AP per class: Average precision computed independently for each defect type
• Precision-recall curve: Plotted by varying confidence thresholds
• [email protected]:0.95: The more stringent metric preferred for high-precision manufacturing use cases

A technique to artificially expand training datasets by applying random but realistic transformations, directly improving YOLO's robustness to factory-floor variability. Mosaic augmentation, introduced in YOLOv4, stitches four images together to improve small-object detection.

• Geometric transforms: Rotation, scaling, translation, and shearing
• Photometric transforms: Brightness, contrast, and hue jitter to simulate lighting changes
• Cutout and MixUp: Random region erasure and image blending to reduce overfitting on specific defect patterns

The deployment of trained YOLO models directly on factory-floor hardware—such as NVIDIA Jetson modules or Intel Neural Compute Sticks—eliminating cloud latency. YOLO's single-stage architecture makes it inherently suitable for edge deployment.

• Inference latency: YOLOv8n achieves < 10 ms on Jetson Orin for 640×640 inputs
• FP16 quantization: Halves model size with negligible accuracy loss for edge deployment
• TensorRT optimization: NVIDIA's inference compiler further accelerates YOLO on edge GPUs