Optimisation of Deep Learning Small-Object Detectors with Novel Explainable Verification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection systems struggle with small objects due to training biases towards larger objects.
• Existing state-of-the-art detectors lack transparency in their decision-making processes, especially for challenging datasets.

Purpose of the Study:

• To develop a methodology for selecting optimal object detectors for specific applications, focusing on small object detection.
• To enhance the explainability and transparency of object detection models using explainable Artificial Intelligence (XAI) techniques.

Main Methods:

• Systematic performance examination of two distinct deep convolutional networks: YOLO V3 (single-stage) and Faster R-CNN (double-stage).
• Analysis of feature extraction layers, anchor boxes, and data augmentation impact, incorporating XAI principles.
• Investigation of model robustness and decision explanation using various techniques.

Main Results:

• Multi-head YOLO V3 detectors with augmented data showed improved performance, even with fewer anchor boxes.
• The proposed WS-Grad and Concat-Grad visualization techniques generate high-resolution heatmaps, enhancing decision transparency.
• New visualization techniques provide more comprehensive insights into detector decision processes compared to existing methods.

Conclusions:

• The proposed machine learning methodology aids in selecting verifiably optimal object detectors for diverse applications.
• WS-Grad and Concat-Grad offer significant improvements in visualizing and understanding the decision-making of object detection models.
• Enhanced transparency and reliability in AI-driven object detection are crucial for complex tasks.