Computer

Vision
-Messar Aya
-Chouchaoui Med El Bachir
-Benamara Ichrak
 Introduction

In computer vision, tasks like object detection, classification, and segmentation

empower computers to analyze and interpret images.
These three techniques play crucial roles. They not only help computers identify
and recognize objects but also understand their context within an image.
As we dive into each method, we'll explore their unique roles, differences, and the
scenarios where each excels.

2
Object Detection
it identifies and locates objects within an image. It not only
recognizes what the objects are but also determines their
positions by drawing bounding boxes around them.

It uses Object Localisation which is finding what

and where a Single object exists in an image Multiple Objects
Your paragraph text
CAT

Classification Classification + Object Detection

Localisation 3
Classification
Image classification is about identifying the main subject in a picture.The objective is to
determine the "main subject" of the image as accurately as possible.

Basic Process

4
Manual Feature Extraction Automatic Feature Extraction

Classification(SVM,Decision Tree..) Integrated Process

time-consuming Efficiency and Accuracy

Older facial recognition

Advanced facial recognition,
systems, basic object detection.
autonomous vehicles, medical imaging.

5
Image Segmentation
a method of processing and analyzing a digital image by dividing it into multiple
parts or segments, so you can process only the important elements instead of
processing the entire picture.

Pixel-Level Definition

highly precise and localized

understanding of the image.

outlines the exact shape of

objects suitable for complex
scenes with overlapping or
irregular shapes.

6
Types

Semantic segmentation
labels each pixel of an
image with a class

instance segmentation
distinguishes between
individual instances of
objects within those
classes

panoptic segmentation
combines both by
labeling all pixels while
also identifying individual
object instances.

7
Object Detection vs. Classification & Segmentation

8
 Pixel-wise
Out Single label or Bounding boxes around
segmentation
put category detected objects
masks
9
 situations for segmentation
Medical Imaging: Precise identification of Autonomous Driving: Recognizing and
tissues, organs, or abnormalities within distinguishing between lanes, pedestrians, and
scans. It’s essential for accurate diagnosis obstacles at a pixel level. Critical for safe
and treatment planning. navigation.

10
 situations for object detection
Retail and Inventory Management:
Automated systems to count and identify
products

11
Models of Object Detection
R-CNN (Region-based Convolutional Neural Networks)
R-CNN: Main Modules
1. Selective Search
Generates 2,000 region proposals
2. Resizing and Feature Extraction
Regions resized to a fixed pre-defined size
Extracts a feature vector from each region proposal
3. Classification with Pre-trained SVM
Classifies region proposals as background or specific object classes

13
 YOLO (You Only Look Once)
it performs real-time detection ,It is famous because it revolutionized object detection
by achieving high-speed, real-time performance without sacrificing much accuracy
key concepts :

Grid cells
The concept of breaking down the images to grid
cells is unique in YOLO

Single Forward Pass

treats object detection as a single regression
problem, predicting bounding boxes and class
probabilities directly from images.

14
Models of Segmentations
 a convolutional network designed for biomedical image
U-NET segmentation

has an encoder-decoder structure with skip connections.

The encoder captures context, while the decoder enables
precise localization. Skip connections help in retaining spatial
information.

15
16
18
Video Surveillance (Security):

Identifying multiple people or objects in real-time video streams for anomaly detection, facial
recognition, or crowd monitoring.

Challenge: Variability in object size, crowding, and

overlapping individuals in busy areas.

19
 Medical Imaging (Healthcare):

Segmenting organs or identifying anomalies (e.g., tumors) within medical images to assist in diagnosis.

Challenge: Objects vary in size and shape, and

boundaries are often unclear.
a subheading

AI-supported mammogram screening increases breast cancer detection by 20%,

study finds

By Jen Christensen, CNN

6 minute read
Published 6:52 PM EDT, Tue August 1, 2023

20
Challenges of Detecting Multiple Objects in a Single Image

Object Overlap and Occlusion High Object VARIABILITY (Crowded

Scenes)

Dealing with Small Objects LIGHTNING CONDITIONS

21
Challenges Related to Multiple Object Detection

Multiple object detection presents several challenges:

Occlusion: Objects can overlap or be partially hidden, making their detection difficult.

Object Variability: Objects can vary in size, shape, color, and texture, complicating their identification.

Lighting Conditions: Variations in lighting can affect object visibility, making detection less reliable.

Processing Speed: In real-time applications like autonomous driving, detection speed is crucial.

Balanced Data: Models may be biased if training data does not fairly represent all object classes.
4o

22
23