Image recognition applications and dataset preparation - DevFest Baghdad 2018
1 like455 views
The document discusses building an image recognition application for identifying Iraqi currency denominations. It covers preparing a dataset with labeled images of each denomination under different conditions, training an object detection model using the Turi Create library, and exporting the model for use in an iOS/Android application. The key steps are preparing a dataset with 300+ labeled photos of each denomination, training an object detection model on the dataset, evaluating the model, and exporting it for use in a mobile application.
![MODEL TRAINING
import turicreate as tc
import os
# Define all images annotations with bounding box details (I am showing only 1)
annotations = tc.SArray([
[{
“label”:”5000ar",
“type":"rectangle",
“coordinates”:{“y":188.5,"x":207,"width":304,"height":152}
}],
… , … …
])](https://image.slidesharecdn.com/imagerecognition-devfest2018-181119065419/85/Image-recognition-applications-and-dataset-preparation-DevFest-Baghdad-2018-32-320.jpg)
![MODEL TRAINING
import turicreate as tc
import os
# Define all images annotations with bounding box details (I am showing only 1)
annotations = tc.SArray([
[{
“label”:”5000ar",
“type":"rectangle",
“coordinates”:{“y":188.5,"x":207,"width":304,"height":152}
}],
… , … …
])
# 1. Load images (Note: you can ignore 'Not a JPEG file' errors)
data = tc.image_analysis.load_images('mr_turi_ic', with_path=True)
data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path)))
data['annotations'] = tc.SArray(data=annotations, dtype=list)](https://image.slidesharecdn.com/imagerecognition-devfest2018-181119065419/85/Image-recognition-applications-and-dataset-preparation-DevFest-Baghdad-2018-33-320.jpg)
![MODEL TRAINING
import turicreate as tc
import os
# Define all images annotations with bounding box details (I am showing only 1)
annotations = tc.SArray([
[{
“label”:”5000ar",
“type":"rectangle",
“coordinates”:{“y":188.5,"x":207,"width":304,"height":152}
}],
… , … …
])
# 1. Load images (Note: you can ignore 'Not a JPEG file' errors)
data = tc.image_analysis.load_images('mr_turi_ic', with_path=True)
data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path)))
data['annotations'] = tc.SArray(data=annotations, dtype=list)
# Make a train-test split
train_data, test_data = data.random_split(0.8)](https://image.slidesharecdn.com/imagerecognition-devfest2018-181119065419/85/Image-recognition-applications-and-dataset-preparation-DevFest-Baghdad-2018-34-320.jpg)
![MODEL TRAINING
import turicreate as tc
import os
# Define all images annotations with bounding box details (I am showing only 1)
annotations = tc.SArray([
[{
“label”:”5000ar",
“type":"rectangle",
“coordinates”:{“y":188.5,"x":207,"width":304,"height":152}
}],
… , … …
])
# 1. Load images (Note: you can ignore 'Not a JPEG file' errors)
data = tc.image_analysis.load_images('mr_turi_ic', with_path=True)
data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path)))
data['annotations'] = tc.SArray(data=annotations, dtype=list)
# Make a train-test split
train_data, test_data = data.random_split(0.8)
# Create a model using Turi Create's object detector API
model = tc.object_detector.create(train_data, max_iterations=1000)
# Save the predictions to an SArray
predictions = model.predict(test_data)
# Evaluate the model and save the results into a dictionary
metrics = model.evaluate(test_data)
print('Precision' , metrics['mean_average_precision'])](https://image.slidesharecdn.com/imagerecognition-devfest2018-181119065419/85/Image-recognition-applications-and-dataset-preparation-DevFest-Baghdad-2018-35-320.jpg)
![MODEL TRAINING
import turicreate as tc
import os
# Define all images annotations with bounding box details (I am showing only 1)
annotations = tc.SArray([
[{
“label”:”5000ar",
“type":"rectangle",
“coordinates”:{“y":188.5,"x":207,"width":304,"height":152}
}],
… , … …
])
# 1. Load images (Note: you can ignore 'Not a JPEG file' errors)
data = tc.image_analysis.load_images('mr_turi_ic', with_path=True)
data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path)))
data['annotations'] = tc.SArray(data=annotations, dtype=list)
# Make a train-test split
train_data, test_data = data.random_split(0.8)
# Create a model using Turi Create's object detector API
model = tc.object_detector.create(train_data, max_iterations=1000)
# Save the predictions to an SArray
predictions = model.predict(test_data)
# Evaluate the model and save the results into a dictionary
metrics = model.evaluate(test_data)
print('Precision' , metrics['mean_average_precision'])
# Save the model for later use in Turi Create
model.save(‘turi_ic.model')
# Export for use in Core ML file to the current directory
model.export_coreml('turi_ic.mlmodel')](https://image.slidesharecdn.com/imagerecognition-devfest2018-181119065419/85/Image-recognition-applications-and-dataset-preparation-DevFest-Baghdad-2018-36-320.jpg)
Image recognition applications and dataset preparation - DevFest Baghdad 2018
- 2. Image Recognition Applications & Dataset Preparation
- 3. Did you see this before ? •Cover violence or nudity images facebook app AppChief
- 4. Did you see this before ? •Suggests sharing photos with recognized facebook friends Moments app AppChief
- 5. Did you see this before ? •Search for photo content in iOS and Android Photos AppChief
- 6. Did you see this before ? •Cutting out a person from an image Sticky app
- 7. Did you see this before ?
- 8. What is image recognition ? How to build my own app ? Why I need it ? Is the ability of software to identify objects, places, people, writing and actions in images. labeling the content of images with meta-tags, performing image content search and guiding autonomous robots, self-driving cars and accident avoidance systems…etc Next slides
- 9. How to build my own app ? TRAIN THE MODEL PREPARE DATASET BUILD AND RUN … …
- 10. What do you need ? PREPARE DATASET Classify image ? Detect multiple objects inside image ? or
- 12. vs OBJECT DETECTION input output Image Image Class labelsClass label OBJECT DETECTION INSTANCE SEGMENTATION types + bounding box + bounding boxes + segmentation PREPARE DATASET Image recognition types - IMAGE CLASSIFICATION CLASSIFICATION CLASSIFICATION + LOCALIZATION
- 13. vs IMAGE CLASSIFICATION OBJECT DETECTION CLASSIFICATION CLASSIFICATION + LOCALIZATION OBJECT DETECTION INSTANCE SEGMENTATION types PREPARE DATASET Image recognition types Example
- 14. vs IMAGE CLASSIFICATION OBJECT DETECTION CLASSIFICATION CLASSIFICATION + LOCALIZATION OBJECT DETECTION INSTANCE SEGMENTATION types PREPARE DATASET Image recognition types WE WILL CONTINUE WITH OBJECT DETECTION
- 15. Let’s create a money reader model STEP 1. Naming objects (Object labels) PREPARE DATASET / OBJECT DETECTION 1 IQD_50000_ar 3 IQD_25000_ar 5 IQD_10000_ar 7 IQD_5000_ar 9 IQD_1000_ar 11 IQD_500_ar 13 IQD_250_ar 2 IQD_50000_en 4 IQD_25000_en 6 IQD_10000_en 8 IQD_5000_en 10 IQD_1000_en 12 IQD_500_en 14 IQD_250_en
- 16. Let’s create a money reader model STEP 2. Take photos for each money face AS MUCH AS YO CAN PREPARE DATASET / OBJECT DETECTION For better accuracy take hundreds of photos with • Different backgrounds • Different positions • Different light conditions • Different orientations
- 17. Let’s create a money reader model PREPARE DATASET / OBJECT DETECTION • Different backgrounds • Different positions • Different light conditions • Different orientations STEP 2. Take photos for each money face AS MUCH AS YO CAN
- 18. STEP 2. Take photos for each money face AS MUCH AS YO CAN Let’s create a money reader model PREPARE DATASET / OBJECT DETECTION
- 19. Let’s create a money reader model 2. Take about 300 photos for each money face PREPARE DATASET / OBJECT DETECTION
- 20. Let’s create a money reader model 2. Take about 300 photos for each money face PREPARE DATASET / OBJECT DETECTION
- 21. Let’s create a money reader model 2. Take about 300 photos for each money face PREPARE DATASET / OBJECT DETECTION
- 22. Let’s create a money reader model STEP 3. Labeling objects inside images PREPARE DATASET / OBJECT DETECTION Label : IQD_50000_en x : 6 y : 120 width : 150 heigh : 370 Objects : Object #1
- 23. Let’s create a money reader model STEP 3. Labeling objects inside images PREPARE DATASET / OBJECT DETECTION Label : IQD_50000_en x : 90 y : 125 width : 313 heigh : 313 Objects : Object #1
- 24. CAUTION : Some training libraries prefers different coordinates system in labeling PREPARE DATASET / OBJECT DETECTION X, Y, Width , Height midX, midY, Width , Height minX, minY, maxX , maxY It’s recommended to check the library needs you want to use for training before start labeling
- 25. How many photos do you think we need for each label ? PREPARE DATASET / OBJECT DETECTION 10 ? 20 ? 50 ? 100 ? 200 ? 300 ?
- 26. Assuming 300 photos is good for our model let’s calculate time required PREPARE DATASET / OBJECT DETECTION 300 image x 14 Label x (5 sec) taking photo x (30 sec) labeling 175 hours !!! 7 Days !
- 27. We made a timesaving app PREPARE DATASET / OBJECT DETECTION Only 1 hour 300 image x 14 Label x (5 sec) taking photo x (60 sec) labeling 50 image x 14 Label x (5 sec) taking photo x (0 sec) labeling
- 28. PREPARE DATASET / OBJECT DETECTION 1 2 3Create labels Capture Generate
- 29. PREPARE DATASET / OBJECT DETECTION 3 Transfer Easy Dataset
- 30. PREPARE DATASET / OBJECT DETECTION Demo time Easy Dataset
- 31. PREPARE DATASET / OBJECT DETECTION Money Reader - العملة قارئ Final live app Final dataset kaggle.com/husamaamer/iraqi-currency- ~1 GB Iraqi Money العراقية العملة Object detection dataset for Iraqi currency
- 32. MODEL TRAINING import turicreate as tc import os # Define all images annotations with bounding box details (I am showing only 1) annotations = tc.SArray([ [{ “label”:”5000ar", “type":"rectangle", “coordinates”:{“y":188.5,"x":207,"width":304,"height":152} }], … , … … ])
- 33. MODEL TRAINING import turicreate as tc import os # Define all images annotations with bounding box details (I am showing only 1) annotations = tc.SArray([ [{ “label”:”5000ar", “type":"rectangle", “coordinates”:{“y":188.5,"x":207,"width":304,"height":152} }], … , … … ]) # 1. Load images (Note: you can ignore 'Not a JPEG file' errors) data = tc.image_analysis.load_images('mr_turi_ic', with_path=True) data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path))) data['annotations'] = tc.SArray(data=annotations, dtype=list)
- 34. MODEL TRAINING import turicreate as tc import os # Define all images annotations with bounding box details (I am showing only 1) annotations = tc.SArray([ [{ “label”:”5000ar", “type":"rectangle", “coordinates”:{“y":188.5,"x":207,"width":304,"height":152} }], … , … … ]) # 1. Load images (Note: you can ignore 'Not a JPEG file' errors) data = tc.image_analysis.load_images('mr_turi_ic', with_path=True) data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path))) data['annotations'] = tc.SArray(data=annotations, dtype=list) # Make a train-test split train_data, test_data = data.random_split(0.8)
- 35. MODEL TRAINING import turicreate as tc import os # Define all images annotations with bounding box details (I am showing only 1) annotations = tc.SArray([ [{ “label”:”5000ar", “type":"rectangle", “coordinates”:{“y":188.5,"x":207,"width":304,"height":152} }], … , … … ]) # 1. Load images (Note: you can ignore 'Not a JPEG file' errors) data = tc.image_analysis.load_images('mr_turi_ic', with_path=True) data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path))) data['annotations'] = tc.SArray(data=annotations, dtype=list) # Make a train-test split train_data, test_data = data.random_split(0.8) # Create a model using Turi Create's object detector API model = tc.object_detector.create(train_data, max_iterations=1000) # Save the predictions to an SArray predictions = model.predict(test_data) # Evaluate the model and save the results into a dictionary metrics = model.evaluate(test_data) print('Precision' , metrics['mean_average_precision'])
- 36. MODEL TRAINING import turicreate as tc import os # Define all images annotations with bounding box details (I am showing only 1) annotations = tc.SArray([ [{ “label”:”5000ar", “type":"rectangle", “coordinates”:{“y":188.5,"x":207,"width":304,"height":152} }], … , … … ]) # 1. Load images (Note: you can ignore 'Not a JPEG file' errors) data = tc.image_analysis.load_images('mr_turi_ic', with_path=True) data['label'] = data['path'].apply(lambda path: os.path.basename(os.path.dirname(path))) data['annotations'] = tc.SArray(data=annotations, dtype=list) # Make a train-test split train_data, test_data = data.random_split(0.8) # Create a model using Turi Create's object detector API model = tc.object_detector.create(train_data, max_iterations=1000) # Save the predictions to an SArray predictions = model.predict(test_data) # Evaluate the model and save the results into a dictionary metrics = model.evaluate(test_data) print('Precision' , metrics['mean_average_precision']) # Save the model for later use in Turi Create model.save(‘turi_ic.model') # Export for use in Core ML file to the current directory model.export_coreml('turi_ic.mlmodel')