0% found this document useful (0 votes)
12 views3 pages

Deep Learning Models

Deep learning

Uploaded by

sudheer Kumar Y
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views3 pages

Deep Learning Models

Deep learning

Uploaded by

sudheer Kumar Y
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 3

1.

Data Collection

 Satellite Images: Use high-resolution satellite images from sources like Sentinel-2 or
Landsat-8.
 Ground Truth Data: Obtain labeled data for different land cover classes, such as urban
areas, water bodies, forests, agricultural fields, etc.

2. Data Preprocessing

 Resizing: Resize images to a fixed size suitable for the U-Net architecture.
 Normalization: Normalize pixel values to the range [0, 1] or [-1, 1].
 Data Augmentation: Apply transformations such as rotations, flips, and zooms to
increase the diversity of the training data.

3. Model Architecture

 Encoder: Apply a series of convolutional layers and max pooling to downsample the
input image and extract features.
 Decoder: Use upsampling and concatenation with corresponding encoder layers to
restore spatial resolution and refine segmentation.

4. Training

 Data Splitting: Divide the dataset into training, validation, and test sets.
 Training: Train the U-Net model using the training set and validate using the validation
set.
 Evaluation: Evaluate the model's performance using metrics like Intersection over Union
(IoU) or pixel accuracy.

5. Prediction and Post-Processing

 Prediction: Use the trained model to predict LULC classes on new satellite images.
 Post-Processing: Apply spatial smoothing or morphological operations to refine the
segmentation results.

6. Visualization

 Visualize the segmented LULC maps to interpret the results and assess the quality of the
segmentation.
Difference Between CNN Architectures, Semantic Segmentation
Networks, and Object Detection Networks
1. CNN Architectures

CNNs (Convolutional Neural Networks) are designed to automatically and adaptively learn
spatial hierarchies of features from input images. They are primarily used for image
classification tasks.

 Key Components:
o Convolutional Layers: Apply filters to input images to produce feature maps.
o Pooling Layers: Reduce the spatial dimensions of the feature maps.
o Fully Connected Layers: Used at the end for final classification.

 Common Architectures:
o LeNet-5: Early architecture for digit recognition.
o AlexNet: Deep network that won ImageNet competition.
o VGGNet: Known for simplicity and depth.
o GoogLeNet (Inception): Introduced Inception modules.
o ResNet: Introduced residual connections to train very deep networks.
o DenseNet: Each layer receives inputs from all previous layers.
o MobileNet: Designed for mobile and embedded applications.
o EfficientNet: Balances depth, width, and resolution for better performance.

2. Semantic Segmentation Networks

Semantic Segmentation involves classifying each pixel in an image to a class, providing dense
predictions and a more granular understanding of the image.

 Key Components:
o Encoder: Extracts features using convolutional and pooling layers.
o Decoder: Upsamples features to original image size, providing pixel-wise classification.

 Common Architectures:
o FCN (Fully Convolutional Network): Replaces fully connected layers with convolutional
layers to produce pixel-wise output.
o U-Net: U-shaped architecture with symmetric encoder and decoder paths, used
extensively in medical imaging.
o SegNet: Encoder-decoder architecture with indices for upsampling in the decoder path.
o DeepLab: Uses atrous convolutions and fully connected CRFs for capturing context at
multiple scales.
o PSPNet (Pyramid Scene Parsing Network): Uses pyramid pooling module to aggregate
context information.
3. Object Detection Networks

Object Detection involves identifying and localizing objects within an image, providing
bounding boxes and class labels for each detected object.

 Key Components:
o Backbone Network: Extracts features from the input image.
o Region Proposal Network (RPN): Generates potential bounding box proposals.
o Detection Head: Classifies and refines bounding box proposals.

 Common Architectures:
o R-CNN (Region-based Convolutional Neural Network): Generates region proposals
using selective search and applies CNN to each proposal.
o Fast R-CNN: Improves R-CNN by using a single-stage training process.
o Faster R-CNN: Introduces RPN to generate proposals directly from feature maps.
o YOLO (You Only Look Once): Single-stage network that directly predicts bounding boxes
and class probabilities from full images.
o SSD (Single Shot MultiBox Detector): Similar to YOLO but uses multiple feature maps at
different scales for detecting objects.
o RetinaNet: Uses a feature pyramid network (FPN) and introduces the focal loss to
handle class imbalance.

Summary

 CNN Architectures: Focus on image classification, identifying the overall class of the image.
 Semantic Segmentation Networks: Provide dense predictions, classifying each pixel in the
image.
 Object Detection Networks: Identify and localize multiple objects within an image, providing
bounding boxes and class labels.

Each type of network has its specific use cases, with CNN architectures providing overall image
classification, semantic segmentation networks offering detailed pixel-level classification, and
object detection networks enabling object localization and classification within images.

You might also like