A CNN Model for Semantic Person Part Segmentation with Capacity Optimization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Optimizing Convolutional Neural Network (CNN) capacity for person part segmentation faces challenges due to limited datasets and reliance on single-modality CNNs.
• Existing methods struggle with effective learning and generalization.

Purpose of the Study:

• To propose a deep learning model with optimal capacity for improved person part segmentation performance.
• To address limitations of previous approaches by leveraging diverse datasets and novel feature enrichment techniques.

Main Methods:

• Utilized external datasets for CNN pre-processing module to enhance image data without extensive manual annotation.
• Developed a novel integration of two complementary modules to enrich feature representations for robust inferences.
• Investigated factors influencing CNN model capacity and proposed two new methods for capacity optimization tailored to task complexity.

Main Results:

• Achieved significant segmentation performance improvement through pre-processing.
• Mitigated over-fitting and under-fitting issues by optimizing CNN capacity.
• Demonstrated superior performance compared to state-of-the-art deep learning models.

Conclusions:

• The proposed deep learning model offers enhanced generalization ability and reduced computational complexity for person part segmentation.
• The novel capacity optimization methods effectively address over-fitting and under-fitting problems.
• This work advances the field of semantic segmentation through improved model design and training strategies.