A Method for Constructing a Loss Function for Multi-Scale Object Detection Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection networks commonly use multiscale pyramid features to identify objects of varying sizes.
• Shallow features typically detect small objects, while deep features detect large objects.
• Current loss functions treat all object samples equally, irrespective of their size and corresponding feature layer, potentially hindering performance.

Purpose of the Study:

• To propose a novel loss function, Predicted Probability Loss (PP-Loss), to enhance object detection performance.
• To address the limitation of existing loss functions in considering the relationship between object size and feature layer representation.
• To improve the accuracy and convergence speed of object detection networks.

Main Methods:

• Developed a Predicted Probability Loss (PP-Loss) function that statistically determines the prediction probability of each feature layer for objects based on label size.
• Integrated PP-Loss into the training process to adjust sample anchor weights, guiding network learning.
• Validated the method on various networks, primarily using YOLO (You Only Look Once) as the core architecture.

Main Results:

• Experimental results demonstrate improved convergence speed during network training.
• The proposed PP-Loss led to enhanced accuracy in object detection tasks.
• Improvements were observed across different network architectures tested.

Conclusions:

• The novel PP-Loss effectively addresses the limitations of traditional loss functions in object detection.
• By incorporating label size-based probabilities, PP-Loss optimizes feature layer utilization for improved detection.
• The method offers a promising approach for enhancing the performance of deep learning-based object detection systems.