Document layout analysis in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Document layout analysis

Problem:You want to teach a computer to recognize different parts of a document page, like titles, paragraphs, images, and tables.

Current Metrics:Training accuracy: 98%, Validation accuracy: 70%, Validation loss: 1.2

Issue:The model is overfitting: it performs very well on training data but poorly on new unseen documents.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 92%.

You can only change the model architecture and training parameters.

You cannot add more training data.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

Computer Vision

import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.callbacks import EarlyStopping

# Sample simplified model for document layout analysis
model = models.Sequential([
    layers.Conv2D(32, (3,3), activation='relu', input_shape=(256, 256, 3)),
    layers.MaxPooling2D(2,2),
    layers.Dropout(0.3),
    layers.Conv2D(64, (3,3), activation='relu'),
    layers.MaxPooling2D(2,2),
    layers.Dropout(0.3),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.4),
    layers.Dense(5, activation='softmax')  # 5 classes: title, paragraph, image, table, other
])

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Early stopping to stop training when validation loss stops improving
early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

# Assuming X_train, y_train, X_val, y_val are prepared image and label datasets
# model.fit(X_train, y_train, epochs=50, batch_size=32, validation_data=(X_val, y_val), callbacks=[early_stop])

Added dropout layers after convolution and dense layers to reduce overfitting.

Lowered learning rate from 0.001 to 0.0005 for smoother training.

Added early stopping callback to stop training when validation loss stops improving.

Results Interpretation

Before: Training accuracy 98%, Validation accuracy 70%, Validation loss 1.2

After: Training accuracy 90%, Validation accuracy 87%, Validation loss 0.6

Adding dropout and early stopping helps the model generalize better by reducing overfitting, improving validation accuracy while slightly lowering training accuracy.

Bonus Experiment

Try using data augmentation techniques like random rotations, zooms, or flips on the training images to further improve validation accuracy.

💡 Hint

Use TensorFlow's ImageDataGenerator or tf.image functions to apply augmentations during training.

Practice

1. What is the main goal of document layout analysis in computer vision?

easy

A. To compress document files for storage

B. To find and label different parts of a document like text, images, and tables

C. To translate documents into different languages

D. To convert handwritten notes into typed text

Solution

1. Step 1: Understand the purpose of document layout analysis

   Document layout analysis is used to detect and label parts of a document such as text blocks, images, and tables.

2. Step 2: Compare options with the purpose

   Only To find and label different parts of a document like text, images, and tables matches this purpose exactly, while others describe different tasks like translation or compression.

3. Final Answer:

   To find and label different parts of a document like text, images, and tables -> Option B

4. Quick Check:

   Document layout analysis = labeling document parts [OK]

Hint: Focus on labeling parts of a page, not translating or compressing [OK]

Common Mistakes:

• Confusing layout analysis with OCR text recognition
• Thinking it translates or compresses documents
• Mixing layout analysis with handwriting recognition

2. Which of the following is the correct way to import Detectron2's layout model in Python?

easy

A. import detectron2.LayoutModel

B. from detectron2 import LayoutModel

C. from detectron2.layout import LayoutModel

D. from detectron2.models import LayoutModel

Solution

1. Step 1: Recall Detectron2 module structure

   Detectron2's layout model is accessed via the 'layout' submodule, so the import should be from detectron2.layout.

2. Step 2: Match options with correct syntax

   from detectron2.layout import LayoutModel is the correct syntax. The other options use incorrect module paths or syntax.

3. Final Answer:

   from detectron2.layout import LayoutModel -> Option C

4. Quick Check:

   Correct import path = from detectron2.layout import LayoutModel [OK]

Hint: Remember submodules come after main package with dot notation [OK]

Common Mistakes:

• Using uppercase import paths incorrectly
• Trying to import directly from detectron2 without submodule
• Using wrong syntax like 'import detectron2.LayoutModel'

3. Given this Python code snippet using Detectron2's layout model:

model = LayoutModel('lp://PubLayNet/faster_rcnn_R_50_FPN_3x/config')
outputs = model.detect(image)
print(len(outputs))

What does len(outputs) represent?

medium

A. The number of classes the model can detect

B. The number of pixels in the input image

C. The number of layers in the model

D. The number of detected layout elements like text blocks and images

Solution

1. Step 1: Understand what model.detect returns

   The detect method returns a list of detected layout elements such as text blocks, tables, and images.

2. Step 2: Interpret len(outputs)

   Taking the length of outputs gives the count of detected elements in the image.

3. Final Answer:

   The number of detected layout elements like text blocks and images -> Option D

4. Quick Check:

   len(outputs) = count of detected elements [OK]

Hint: Outputs list length = number of detected layout parts [OK]

Common Mistakes:

• Thinking it counts pixels or model layers
• Confusing output length with number of classes
• Assuming outputs is a single prediction, not a list

4. You wrote this code to detect layout elements but get an error:

model = LayoutModel('lp://PubLayNet/faster_rcnn_R_50_FPN_3x/config')
outputs = model.detect()
print(outputs)

What is the likely cause of the error?

medium

A. The detect method requires an image argument but none was given

B. The model path is incorrect

C. The print statement syntax is wrong

D. LayoutModel cannot be instantiated without extra parameters

Solution

1. Step 1: Check method usage

   The detect method requires an input image to analyze, but the code calls detect() without any argument.

2. Step 2: Identify error cause

   Missing the required image argument causes a TypeError or similar error.

3. Final Answer:

   The detect method requires an image argument but none was given -> Option A

4. Quick Check:

   detect() needs image input [OK]

Hint: Always pass the image to detect() method [OK]

Common Mistakes:

• Forgetting to pass the image to detect()
• Assuming model path is wrong without checking error
• Thinking print syntax causes error

5. You want to improve document layout analysis accuracy on scanned forms with many tables. Which approach is best?

hard

A. Fine-tune a Detectron2 layout model on a labeled dataset of scanned forms

B. Use a generic OCR tool without layout detection

C. Increase image resolution without changing the model

D. Manually draw bounding boxes on each form

Solution

1. Step 1: Identify the goal

   The goal is to improve accuracy specifically for scanned forms with many tables.

2. Step 2: Evaluate options for improving accuracy

   Fine-tuning a layout model on a relevant labeled dataset adapts it to the specific document type, improving accuracy. Generic OCR ignores layout. Increasing resolution alone may not help. Manual bounding boxes are not scalable.

3. Final Answer:

   Fine-tune a Detectron2 layout model on a labeled dataset of scanned forms -> Option A

4. Quick Check:

   Fine-tuning on target data = best accuracy boost [OK]

Hint: Train model on similar documents for best results [OK]

Common Mistakes:

• Relying only on OCR without layout context
• Thinking higher resolution fixes layout detection
• Ignoring the need for labeled data to fine-tune