A Dual Neural Architecture Combined SqueezeNet with OctConv for LiDAR Data Classification

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Light Detection and Ranging (LiDAR) is a crucial technique for data acquisition with broad applications.
• Deep Convolutional Neural Networks (CNNs) show promise for classifying LiDAR-derived Digital Surface Models (LiDSM).
• Existing CNNs often suffer from excessive parameters and spatial redundancy, limiting efficiency.

Purpose of the Study:

• To develop a more accurate and efficient deep learning model for LiDAR-DSM data classification.
• To address the limitations of parameter count and spatial redundancy in traditional CNNs.

Main Methods:

• Proposed a novel dual neural architecture named OctSqueezeNet.
• Integrated SqueezeNet's Fire modules (using 1x1 convolutions) into Octave Convolution (OctConv) layers.
• OctConv reduces spatial redundancy by processing feature maps at different resolutions.

Main Results:

• OctSqueezeNet demonstrated competitive advantages in classification accuracy.
• The proposed method significantly reduced the computational amount compared to state-of-the-art methods.
• Experiments were validated on two well-known LiDAR datasets.

Conclusions:

• OctSqueezeNet offers an effective solution for improving both accuracy and efficiency in LiDAR data classification.
• The hybrid architecture successfully mitigates parameter bloat and spatial redundancy issues in deep learning models.