Computer Vision Masterclass: Build Advanced CNNs

Explore how convolutional neural networks process images by drawing analogies to the visual cortex, from simple orientation features to complex object recognition across cortical layers V1 to V6.

Convolution functions provide a linear operation to extract features in images across 1d, 2d, and 3d, enabling CNNs to train filters automatically from data.

Study how convolutional layers use filter size, stride, and zero padding to shape output volumes, and how pooling and parameter sharing curb complexity and overfitting.

Explore the evolution of cnn architectures from LeNet-5 to AlexNet, ZFNet, and Gee-Gee net, highlighting convolution, pooling patterns, activation, data augmentation, dropout, and visualization techniques.

Explore CNN architectures from Inception to residual networks, learning how multi-path filters and skip connections enable deeper models with fewer parameters. Observe how average pooling replaces fully connected layers.

Learn how CNNs perform image localization, detection, segmentation, and captioning-related tasks like style transfer by predicting bounding boxes with regression or classification, using sliding-window and multi-scale approaches.

Examine how object detection extends localization with region proposals and CNNs, tracing R-CNN variants, region proposal networks, and fast end-to-end approaches like YOLO and faster R-CNN.

Explore image segmentation, covering semantic segmentation with pixel-level labeling and instance segmentation from region proposals. Learn about learnable upsampling, skip connections, and encoder-decoder architectures that drive end-to-end segmentation.

Explore deep dream and style transfer visuals to gain intuition into cnn operation, and review visualization methods from weight analysis to deconvolution and criterion-based optimization.

Explore deep dream, a backpropagation based technique that amplifies activations in the inception 4C layer. Apply Noorul style transfer with content and style losses, gram matrices, and patches.

Develops super resolution for image enhancement via supervised learning and regenerative adversarial networks, restoring details and upscaling images for HD content, microscopy, and forensics, measured by PSNR and opinion score.

Explore generative adversarial networks (gan), including dcgan, with a generator and discriminator that synthesize realistic images from noise, enabling unsupervised learning and super-resolution techniques like srgan, while noting training challenges.

Explore the biological basis of CNNs and the neurophysiology of vision across the visual cortex, review CNN architectures with layers and spatial arrangements, and master convolution, deconvolution, stride, and padding.