Neural Inference Search for Multiloss Segmentation Models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semantic segmentation is crucial for surveillance but current models lack required tolerance in complex environments.
• Existing deep learning segmentation models struggle with multi-class tasks and varied environmental conditions.

Purpose of the Study:

• To introduce a novel algorithm, Neural Inference Search (NIS), for optimizing deep learning semantic segmentation models.
• To enhance model performance through simultaneous optimization of learning and multiloss parameters.

Main Methods:

• Developed Neural Inference Search (NIS) algorithm for hyperparameter optimization.
• Incorporated three novel search behaviors: Maximized Standard Deviation Velocity Prediction, Local Best Velocity Prediction, and n-dimensional Whirlpool Search.
• Utilized a scheduling mechanism to manage search behavior contributions and employed Long Short-Term Memory (LSTM)-Convolutional Neural Network (CNN) for predictions.

Main Results:

• NIS-optimized models demonstrated significant improvements in multiple performance metrics across five segmentation datasets.
• Achieved superior results compared to state-of-the-art segmentation methods and models optimized with other search algorithms.
• NIS consistently yielded better solutions for numerical benchmark functions than various search methods.

Conclusions:

• Neural Inference Search (NIS) offers a robust and effective approach to optimizing deep learning semantic segmentation models.
• The proposed algorithm significantly advances the capabilities of semantic segmentation for demanding surveillance applications.
• NIS provides a reliable method for improving model performance and solving complex optimization problems.