[NEW] 2026:Mastering Computer Vision With GenAI :12 Projects

Explore the intuition behind neural networks, from logistic regression to multi-layer architectures with input, hidden, and output layers, using weighted features, biases, and activation functions to compute probabilities.

Clarify the problem goal and data type, then assess data quality and size to decide between machine learning and deep learning, considering latency and available pre-trained models.

Discover the five steps of the deep learning model life cycle using tf.keras, including defining the model, compiling, fitting, evaluating, and making predictions with sequential or functional APIs.

Explore TensorFlow Keras model design by comparing sequential and functional APIs, building with dense layers and input shapes, and understanding trade-offs between quick prototypes and flexible architectures.

Build a multilayer perceptron with the Keras sequential API for mnist digits. Load data one hot encode normalize reshape to 784 features and train with Adam using categorical cross entropy.

Learn to build a Keras functional API model for fashion MNIST, including data loading, scaling, one-hot encoding, and dense layers with dropout and batch normalization.

Explore cost functions and gradient descent in linear regression, minimizing prediction errors with beta values and weights, while learning to prevent underfitting and overfitting through cross validation.

Improve model performance by using Adam, compare with SGD or rmsprop, and apply batch normalization with dropout, activation functions, and L1/L2 regularization to stabilize gradients.

Learn to build a feedforward neural network in Keras for binary classification, including data standardization, batch normalization, dropout, regularization, class weights, and callbacks for checkpoints and early stopping.

Explore gradient descent and its optimizers, including batch, stochastic, and mini-batch types, along with momentum, Rmsprop, and Adam, to understand navigating local and global minima with effective learning rates.

Learn keras loss functions for regression and classification, including mean squared error, mean absolute error, binary cross entropy, hinge losses, and multiclass and sparse cross entropy, plus KL divergence.

Develop a cnn on the cifar-10 dataset using tf keras, with conv2d, maxpool2d, flatten, and dense layers to classify 32 by 32 by 3 rgb images into ten classes.

Train a model to classify numbers on the mNIST dataset using Keras, choosing the sequential or functional API, applying early stopping and checkpoints, reload the model, then save and predict.

Learn to build a CNN classifier for MNIST using both sequential and functional API, including data prep, one-hot encoding, normalization, Conv2D, MaxPooling, flatten, dense, dropout, and softmax activation.

Explore a cnn-based fashion mnist classifier, detailing data reshaping, one-hot encoding, conv and pooling layers, dense layers, training, accuracy, and model saving and deployment steps.

Explore keras preprocessing layers for image, text, and numeric data, including image resizing, rescaling, center cropping, and data augmentation with random flip, random translation and zoom, and random rotation.

Apply Keras preprocessing layers to a cats and dogs dataset, demonstrating resize, random augmentations (height, width, zoom, flip, rotate, crop, translation, brightness, contrast) and Google Drive directory based loading.

Apply transfer learning on a cats and dogs dataset using a pre-trained MobileNet V2, implement data augmentation with image data generators, and build a two-class classifier in Colab.

Mount Google Drive, import packages, and prepare a two-branch transfer learning model by concatenating MobileNet v2 and DenseNet 169 for pneumonia vs normal chest x-ray classification.

Explore LSTM and GRU architectures, detailing cell state, gates, and how information is retained across sequences, compare training time and parameters, and illustrate many-to-many and autocomplete examples.

Train a generative adversarial network on Fashion-MNIST using a generator and discriminator, implemented in TensorFlow Keras, trained on Google Colab with random noise input and image scaling.

Clean and encode image captions by filtering poorly encoded images, building a vocabulary with a frequency threshold, and aligning captions with a glove-based embedding for an image captioning model.

Develop an image captioning model by combining encoded image features and caption input through dropout, embedding, and lstm decoders, then train with cross-entropy loss and softmax to generate captions.

Compare semantic and instance segmentation, showing how CNNs assign pixel-level labels to categories and distinguish individual object instances for scene understanding and autonomous driving.

Explore four segmentation algorithms—region based, edge detection, clustering, and Mask R-CNN—and learn how semantic and instance segmentation, object detection, and localization identify and delineate objects.

Explore two-step object detection by combining region proposals with classification and localization, and illustrate how segmentation and bounding boxes identify and count multiple object instances in images.

Describe fast r-cnn, an upgrade from sbp net using roi pooling on a single convnet, with softmax classifier and linear bounding-box regression in a joint loss.

Explore mask r-cnn, an extension of faster r-cnn that integrates pixel-wise masks during training for per-pixel classification and object segmentation. Its lower branch provides segmentation for each bounding box.

Explore how SSD uses a base network (VGG16 or ResNet) with extra layers to form multi-scale maps and predict scores and box offsets for 8732 default boxes.

Explore image annotation tools like make sense and Roboflow to build datasets for object detection and recognition, using polygons or bounding boxes and exporting to COCO, VOC XML, or YOLO.