Appearance

AI Model Types ​

What are AI Models? ​

Artificial intelligence models are algorithms trained to recognize patterns in images and make automatic predictions. Rosepetal supports different types of models according to specific visual inspection needs.

Available Model Types ​

🏷️ Classification (MULTICLASS) ​

What does it do? ​

Categorizes complete images into different predefined classes or categories.

Use Cases ​

  • Quality control: "Good" vs "Defective"
  • Product categorization: Type A, B, C
  • Defect classification: Scratch, dent, stain
  • Compliance verification: Pass/Fail

Features ​

  • One label per image: Each image belongs to a single class
  • Output with confidence: Percentage of certainty for each prediction
  • Multiple classes: Can handle 2 to N different categories
  • Fast and efficient: Ideal for binary or multiple decisions

Practical Example ​

Screw image → Model → "Good" (95% confidence)
Part image → Model → "Defective" (87% confidence)

🚨 Anomaly Detection (ANOMALY) ​

What does it do? ​

Identifies anomalous elements or out of the ordinary by comparing with learned patterns from "correct" samples.

Use Cases ​

  • Detection of unknown defects: Without previous examples of the defect
  • Surface inspection: Cracks, scratches, deformations
  • Uniformity control: Variations in texture or shape
  • Contamination detection: Unexpected particles or stains

Features ​

  • Unsupervised learning: Only needs "normal" examples
  • Novelty detection: Finds patterns not seen before
  • Anomaly score: Degree of "rarity" of each image
  • Configurable sensitivity: Adjust detection level

Practical Example ​

Train with: 1000 images of perfect parts
Predicts: New image → "Anomaly detected" (marked area)

🎯 Object Detection (imageObjectDetection) ​

What does it do? ​

Locates and classifies multiple objects within the same image, drawing bounding boxes around each one.

Use Cases ​

  • Automatic counting: Number of parts in a tray
  • Component localization: Find screws, chips, connectors
  • Assembly inspection: Verify presence and position of parts
  • Layout analysis: Distribution of elements on surfaces

Features ​

  • Multiple detections: Several objects per image
  • Precise localization: x,y,width,height coordinates of each object
  • Simultaneous classification: Type and location in a single prediction
  • Confidence per object: Individual certainty for each detection

Practical Example ​

Circuit image → Model → [
  "Resistor" (x:100, y:50, confidence: 92%),
  "Capacitor" (x:200, y:80, confidence: 88%),
  "Chip" (x:150, y:120, confidence: 95%)
]

🎨 Segmentation (MULTILABEL) ​

What does it do? ​

Delimits precise areas pixel by pixel of different elements or regions of interest in the image.

Use Cases ​

  • Area measurement: Exact size of defects or components
  • Morphological analysis: Shape, perimeter, area of objects
  • Material separation: Different types of surface
  • Region mapping: Specific areas for detailed analysis

Features ​

  • Pixel precision: Exact boundary delimitation
  • Multiple regions: Different areas in the same image
  • Quantitative analysis: Area, shape, distance measurements
  • Color masks: Color visualization of each region

Practical Example ​

Welding image → Model → Mask that marks:
- Green area: Correct welding (1,234 pixels)
- Red area: Porosity defect (89 pixels)
- Blue area: Base metal (5,678 pixels)

Model Architectures ​

🎓 Supervised ​

Require labeled data for training

  • Classification: Images with class labels
  • Detection: Images with bounding boxes
  • Segmentation: Images with pixel-to-pixel masks

🔍 Unsupervised ​

Learn patterns without explicit labels

  • Anomaly detection: Only "normal" examples
  • Clustering: Automatic grouping by similarity
  • Autoencoders: Image reconstruction

🎯 Semi-supervised ​

Combine labeled and unlabeled data

  • Few examples: When there are few labeled images
  • Active learning: Model suggests what to label
  • Transfer learning: Leverages pre-trained models

Appropriate Model Selection ​

🎯 According to Objective ​

For Binary Decisions ​

  • Recommended: MULTICLASS Classification
  • Alternative: Anomaly detection (if only "good" examples exist)

To Locate Elements ​

  • Recommended: Object detection
  • When: You need to know WHERE the object/defect is

For Precise Measurements ​

  • Recommended: MULTILABEL Segmentation
  • When: You need exact area, shape or perimeter

For Unknown Defects ​

  • Recommended: Anomaly detection
  • When: You don't have examples of all possible defects

📊 According to Available Data ​

With Many Labeled Images ​

  • Classification or Object detection
  • Recommended minimum: 100-500 images per class

With Few Labeled Images ​

  • Transfer learning with pre-trained models
  • Semi-supervised combining with unlabeled images

Only "Normal" Images ​

  • Anomaly detection
  • Ideal when it's difficult to obtain defect examples

⚡ According to Speed Requirements ​

Real Time (< 100ms) ​

  • Classification: Fastest
  • Simple detection: With few objects

Batch Processing ​

  • Segmentation: Detailed analysis
  • Complex detection: With many objects

Evaluation Metrics ​

🏷️ For Classification ​

  • Accuracy: Percentage of total hits
  • Precision per class: Hits for each category
  • Recall: Ability to find positive cases
  • F1-Score: Balance between precision and recall

🎯 For Object Detection ​

  • mAP (mean Average Precision): Average precision per class
  • IoU (Intersection over Union): Box overlap
  • Localization precision: Coordinate accuracy
  • Detection rate: Percentage of objects found

🚨 For Anomaly Detection ​

  • AUC-ROC: Area under the ROC curve
  • False positive rate: Normals classified as anomalous
  • False negative rate: Undetected anomalies
  • Optimal threshold: Balance point between error types

🎨 For Segmentation ​

  • Pixel IoU: Pixel-to-pixel overlap
  • Dice Coefficient: Similarity between masks
  • Edge precision: Accuracy in contours
  • Region completeness: Percentage of correctly segmented area

Implementation Considerations ​

💻 Computational Resources ​

  • GPU recommended: For complex training
  • RAM memory: Varies according to image size and model
  • Storage: For large datasets and model history

🔄 Model Updates ​

  • Periodic retraining: With new data
  • Versioning: Maintain different versions
  • A/B Testing: Compare performance between versions
  • Rollback: Ability to return to previous version