A Review on Multiscale-Deep-Learning Applications.

Area of Science:

• Computer Vision and Deep Learning
• Artificial Intelligence

Background:

• Existing convolutional neural network (CNN) models often lose spatial information and struggle with feature representation due to limitations in capturing multiscale context and semantic information during pooling operations.
• CNNs extract simple semantic features (edges, corners) in early layers and complex features (geometric shapes) in later layers. Utilizing both is crucial for tasks like classification and segmentation.
• Multiscale capability is essential for optimally capturing features at various scales, enabling the fusion of low-level and high-level features to enhance deep model performance.

Purpose of the Study:

• To present a comprehensive taxonomy of multiscale deep learning methods, detailing architectures and their strengths.
• To address the limitations of existing CNNs in handling spatial-information loss and inadequate feature representation.
• To categorize multiscale approaches into feature learning and feature fusion for improved deep learning models.

Main Methods:

• Categorization of multiscale deep learning approaches into two primary types: multiscale feature learning and multiscale feature fusion.
• Multiscale feature learning involves using kernels of various sizes to extract a wider range of features and predict spatial mappings.
• Multiscale feature fusion utilizes features of different resolutions to identify patterns across short and long distances without requiring deep networks.

Main Results:

• The review provides a novel taxonomy classifying multiscale deep learning methods.
• Identified strengths of various implemented architectures within the multiscale deep learning domain.
• Discussed applications of these multiscale techniques in diverse fields such as satellite imagery, medical imaging, agriculture, and industrial systems.

Conclusions:

• Multiscale deep learning methods offer a promising solution to overcome spatial-information loss and enhance feature representation in CNNs.
• The proposed taxonomy provides a structured overview of existing multiscale strategies, aiding researchers in selecting appropriate methods.
• Effective implementation of multiscale feature learning and fusion can significantly boost deep model performance across various real-world applications.