Object Detection and Classification Approaches

Literature Review and Comparative Study
 

Table of Contents

•   What is Object Detection
•  Modern Deep Learning Approaches
•  Object Detection in Side-Scan Sonar (SSS)
•  Comparison of Methods for SSS
•  Evaluation Metrics
•  Preliminary Results
•  Future Work

What is Object Detection 

• Object detection and classification are fundamental tasks in computer vision, used in applications like facial recognition, autonomous driving, medical imaging, and defense systems.
•  Object Detection: Identifies objects in images, providing bounding boxes.
• Classification: Assigns labels to detected objects.

Modern Deep Learning Approaches

• Modern approaches leverage neural networks for improved accuracy and scalability.
• Single-Shot Detection (SSD) vs Dual-Shot Detection.
• YOLO (You Only Look Once): Ideal for high-speed detection, though less effective for smaller objects.
• Region-Based CNNs (R-CNN):High accuracy but slower. Faster R-CNN introduces region proposal networks for better speed.
• Mask R-CNN: Adds segmentation capabilities, enhancing shape analysis at a higher computational cost.

Object Detection in SSS

• SSS images present unique challenges due to noise, low resolution, and texture variability.
• Challenges: High noise levels, shape and texture similarities, and limited annotated data.

Comparison of Methods for SSS

• Real-Time Performance: YOLO excels in speed, suitable for real-time detection, while Faster R-CNN is better for accuracy in complex scenes.
• Multi-Object Detection: Faster R-CNN handle cluttered scenes effectively.
• Noise Adaptability: Domain adaptation and fine tuning techniques are essential for SSS, addressing noise and data scarcity.

•  Classification Metrics:
  • Accuracy:Proportion of correct predictions.
  • Precision & Recall: Balance false positives and false negatives.
  • F1 Score: Harmonic mean of precision and recall.
•  Object Detection Metrics:
  •   IoU (Intersection over Union): Measures overlap between predicted and true bounding boxes.
  • mAP (Mean Average Precision): Precision averaged across classes and IoU thresholds.
  •  Detection Speed: Frames per second (FPS), crucial for real-time tasks.

• The dataset contains 1170 side-scan sonar images collected using a 900–1800 kHz Marine Sonic dual frequency side-scan sonar of a Teledyne Marine Gavia Autonomous Underwater Vehicle (AUV) .
• All the images were carefully analyzed and annotated, including the image coordinates of the Bounding Box (BB) of the detected objects divided into NOn-Mine-like BOttom Objects (NOMBO) and MIne-Like COntacts (MILCO) classes

Dataset

Table 1. Summary of the dataset.

Dataset

Non mine image

Mine image

Preliminary Results

• Initial experiments on sonar datasets reveal key insights:
• Dataset Details: 825 sonar images split into 80% training and 20% testing.
• Classification Results:
  • Before Training: Accuracy of 0.27, indicating poor performance.
  •  After Training: Improved accuracy (0.91) with a ResNet-based classifier.
•  Detection Challenges: Pretrained YOLOv8 struggled with noisy SSS data, emphasizing the need for domain adaptation and fine-tuning.

Future Work

• Future efforts aim to address current limitations:
•  Enhanced Datasets: Include diverse, annotated sonar images to improve training.
• Synthetic Data Generation:Augment datasets with realistic synthetic images.
• Advanced Models: Explore fine-tuned YOLO and R-CNN variants for SSS data.
• These steps will enhance model reliability and accuracy in real-world applications.

THANK YOU