Project - Ship Detection from Satellite Imagery using ML

Introduces run length encoding and decoding concepts for preprocessing ship-detection data from satellite imagery, showing how to convert encodings into a segmented mask and build a reusable function.

Explore run length encoding and decoding as a lossless compression technique, illustrated with examples and row-wise image encoding to reduce data size.

Explore run-length encodings in the dataset and learn to convert encoded pixels into masks for ships, using start pixels and run lengths.

Develop a reusable function to convert encoded pixel runs (RLE) into segmentation masks, enabling automatic generation of training masks from satellite ship images for model learning.

Congratulations on completing the second module; learn to handle diverse data sets, generate segmentation masks with a flexible function, and prepare training, validation data, and preprocessing in the next module.

Prepare the training and validation datasets for ship detection from satellite imagery, covering data preprocessing and augmentation techniques in this module.

Prepare training and validation data for ship detection by cleaning files, computing sizes, encoding shapes, grouping by image id, and applying a stratified train-test split with ship count visualization.

Balance the dataset for ship detection by applying random undersampling across grouped ship counts, using train data frame operations, lambda mapping, and random sampling to achieve even distribution.

Define and document model and training parameters, including batch size, epochs, edge crop, gaussian noise, upsampling mode, net scaling, and image scaling, to prepare training and validation batches.

Create training and validation datasets using a batch generator that yields image and mask pairs for ship detection from satellite imagery, with on-the-fly visualization.

Augment images and masks with a Keras image data generator, applying rotations and flips to expand the dataset and show montages of augmented images and masks with ships bounding boxes.

Learn how garbage collection manages memory by reclaiming unused objects with gc.collect. The video shows how restarting a notebook reduces garbage, illustrating memory management for large data sets.

We completed data preparation and data pre-processing, freeing memory to support training. Congratulations on completing this third module; the training part will be covered in the next module.

Explore the final module of ship detection from satellite imagery using ml, covering cnn architecture, convolutions and pooling layers, and building, compiling with loss and optimizer, plus callbacks and training.

Explore the high-level ideas of CNNs and U-Net for image segmentation, and walk through model building, training, and essential layers like convolution and max pooling.

Explore how convolutional neural networks' convolution layers reduce spatial size while expanding feature maps, apply padding and stride, use convolution 2D transpose and upsampling, and compare max and average pooling.

Explain why UNET and CNN layers like convolution and max pooling suit image segmentation, and how training uses the dice coefficient with dice loss and binary cross entropy.

Understand and build a U-Net for ship detection from satellite images. Discover encoder-decoder structure, skip connections, upsampling and transpose convolutions, and padding effects in a fully convolutional network.

Learn to compile a segmentation model using a dice coefficient metric with smoothing, a combo loss (dice loss plus binary cross-entropy), and an Adam optimizer with learning rate and decay.

Configure callbacks to monitor training performance, saving best weights with model checkpoint, adjusting learning rate on plateaus, and stopping early when no improvement occurs.

Train the segmented ship-detection model with augmented data using fit_generator, compute loss, and save the trained weights for evaluation on test images, while tweaking parameters to boost performance.

Celebrate completing the final module by reviewing unit architecture with convolutional neural network layers, applying the combo loss solution, training your model, and tuning parameters to optimize performance.

Explore convolutional neural networks and the U-Net architecture for image segmentation, from data exploration to run length encoding and preprocessing. Explore Udemy programs and Spark Ignition live trainings.