Image Filtering

بص على محاضرة ال Spatial Filtering في الايميج ارحم

the image is what matters here

حابب أشارك معاكم strategy note عملتها لل2D Convolution، علشان لاحظت اني لما بحاول أعمل 2D Convolution بإيدي على ورقة بتوه جدًا وسط الأرقام، إن شاء الله تفيد حد =)
Performing 2D Convolution

By: Asser Ahmed

1. Machine Learning vs. Deep Learning

Deep Learning (DL) is a highly effective subfield of Machine Learning (ML) that uses multiple layers to learn data representations and find complex patterns.

The fundamental difference between the two lies in Feature Extraction:

• Traditional Machine Learning: The pipeline requires human intervention to extract features. The flow is: Input Data $\to$ Manual Feature Extraction $\to$ Classification Model $\to$ Output (e.g., "Car" or "Not Car"). This can be highly time-consuming and limited by human understanding of the feature dimensions.

• Deep Learning: The network figures out the features on its own. The flow seamlessly merges the steps: Input Data $\to$ Feature Extraction + Classification (handled internally by the network layers) $\to$ Output.

2. Convolutional Neural Networks (CNNs)

When dealing with images (like a 32x32x3 RGB image), standard neural networks struggle because they flatten the image, losing the spatial relationship between pixels.

CNNs are designed specifically to solve this. They preserve the spatial structure of the image and use far fewer weights by sharing them across the entire image.

A standard CNN is a sequence of distinct layers:

A. The Convolution Layer (CONV)

This is the core building block of a CNN. Instead of connecting every pixel to a neuron, a CONV layer uses filters (also called kernels).

• The Sliding Process: If you have a 32x32x3 image and a 5x5x3 filter, the network "slides" (convolves) this filter spatially over the entire image. At every location, it computes the dot product between the filter's weights and the image's pixels.

The Kernel should have the same depth of the previous layer

• Weight Sharing: The network ensures that all locations in the image are processed using the exact same weights from that filter.

• Activation Maps: Sliding one filter across the image produces a single 2D "activation map" (e.g., a 28x28x1 map). If your CONV layer uses 6 different filters, it will produce 6 separate activation maps stacked together (a 28x28x6 volume).

B. Activation Functions (ReLU)

Between the Convolution layers, the network intersperses activation functions (like ReLU) to introduce non-linearity, allowing the network to learn complex, non-linear patterns.

C. The Pooling Layer (POOL)

Pooling layers are used to downsample the activation maps, reducing the spatial dimensions (width and height) while keeping the most important information.

• Max Pooling: The most common technique. It looks at a single depth slice of the activation map and extracts the maximum value within a specific window.

• Example: If you use a 2x2 max pooling filter with a stride of 2 on a 4x4 matrix, it will divide the matrix into four 2x2 blocks and only keep the highest number from each block, resulting in a compacted 2x2 matrix.

D. The Fully Connected Layer (FC)

After the image has passed through multiple sequences of CONV, ReLU, and POOL layers, the high-level features are finally flattened and passed into Fully Connected layers. These act just like a standard neural network to make the final classification (e.g., predicting if the image is a car, truck, airplane, ship, or horse).

more here