Page 219 - Ai_C10_Flipbook
P. 219
CARGO SHIP
STEAM BOAT
CRUISE Cargo Ship 17%
Steam Boat 8%
FISHING BOAT Cruise 75%
INPUT CONVOLUTION - ReLU POOLING CONVOLUTION - ReLU POOLING FLATTEN FULLY SOFTMAX
CONNECTED
Output Probability
FEATURE LEARNING CLASSIFICATION
Input Image
Image Processed by CNN
In the above diagram, an input image is provided, processed through a CNN, and a prediction is generated based
on the labels in the corresponding dataset.
Layers of Convolutional Neural Network (CNN)
The different layers of a Convolutional Neural Network (CNN) are shown in the following figure:
CARGO SHIP
STEAM BOAT
CRUISE
FISHING BOAT
INPUT CONVOLUTION - ReLU POOLING CONVOLUTION - ReLU POOLING FLATTEN FULLY SOFTMAX
CONNECTED
FEATURE LEARNING CLASSIFICATION
A convolutional neural network consists of the following layers:
• Convolution Layer • Rectified Linear Unit (ReLU) • Pooling Layer • Fully Connected Layer
Convolution Layer
The Convolutional Layer is the first layer in a Convolutional Neural Network (CNN) and plays a crucial role in
processing visual data, such as images. Its main objective is to extract key features from the input image, starting
with low-level features like edges, textures, colours, and gradients. These features serve as the building blocks for
the network to understand the content of the image.
In a CNN, this layer is not limited to just one convolutional operation. As the network deepens, additional
convolutional layers are added, each progressively capturing more complex patterns or high-level features
such as shapes, objects, and contextual patterns. This enables the CNN to evolve, recognising increasingly
sophisticated and abstract features, ultimately allowing the network to understand entire images.
The convolution operation applies multiple kernels to an image to extract various features from the image. The
result of this process is called a feature map (or activation map), which highlights the important features detected
by the kernels. The feature map has several key functions:
• Image Size Reduction: It reduces the image size, often through pooling layers that follow convolutional layers,
making it easier and faster to process while retaining crucial information.
• Focus on Relevant Features: It helps focus on the most important features needed for further image processing.
For example, in plant disease detection, the model may only need to focus on the patterns on the leaves (such
as discolouration or spots) rather than analysing the entire plant. By emphasising the specific features of the
leaves, the model can efficiently and accurately detect diseases like fungal infections or bacterial blight, even in
a crowded field of plants.
Computer Vision 217

