Rotation Normalization for DCNN Object Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep Learning systems face challenges in accurately detecting objects in geospatial images due to variations in scale, lighting, and perspective, requiring large numbers of training examples, especially negative training examples, which is inefficient and reduces detection accuracy.
Innovation Solution
The system employs a trainable feature extractor using Deep Convolutional Neural Networks (DCNNs) with rotation normalization and scale normalization, reducing the need for extensive training examples by utilizing the rotation variant property of objects, allowing for near human-level object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large numbers of positive and negative training examples are used to train Deep Learning systems for geospatial image analysis, then object detection accuracy is improved, but the complexity and resource requirements of the training process increase significantly
Solution Approach 1:
The patent extracts and eliminates the need for negative training examples from the training process. By using rotation variant object detection, the system only requires positive training examples (images containing the target object), removing the complex requirement for collecting, curating, and processing large numbers of negative examples while maintaining high detection accuracy.
Solution Approach 2:
The patent applies rotation normalization as a preliminary action before training. By pre-processing training examples to normalize rotation angles and generate rotation variants, the system prepares the data in advance to reduce the need for extensive negative examples during the training phase, thereby simplifying the overall training process.
2Reliability
If extensive negative training examples are provided to teach the system what objects are NOT present, then false positive detection is reduced, but the training time and computational resources required increase significantly
Solution Approach 1:
The patent removes the requirement for negative training examples from the training process. By leveraging rotation variant detection and normalization, the system achieves reliable false positive reduction using only positive training examples, eliminating the time-consuming process of collecting and training on negative examples.
Solution Approach 2:
The patent generates multiple rotation variants of positive training examples to create a comprehensive training set. By copying and rotating the same positive examples at different angles, the system learns to recognize objects regardless of orientation without needing negative examples, thereby reducing training time while maintaining reliability.
3Adaptability or versatility
If the system is trained to recognize objects at multiple scales, lighting conditions, and perspectives, then detection versatility is improved, but the number of training examples required increases exponentially
Solution Approach 1:
The patent applies scale normalization and rotation normalization as preliminary actions to training examples. By pre-processing images to standardize scale and orientation, the system learns versatile detection capabilities from a smaller set of normalized training examples, avoiding the exponential increase in data requirements that would otherwise be needed to cover all scale and perspective variations.
Solution Approach 2:
The patent creates a universal detection approach by training on normalized rotation variants. A single set of normalized positive training examples serves multiple functions, enabling the system to detect objects at various scales, orientations, and perspectives without requiring separate training datasets for each condition, thereby maintaining versatility while reducing the quantity of training examples needed.
Data Source
AI summary
System and method for detecting objects in geospatial images, 3D point clouds and Digital Surface Models (DSMs). Deep Convolution Neural Networks (DCNNs) are trained using positive and negative training examples. Using a rotation pattern match of only positive examples reduces the number of negative examples required. In DCNNs softmax probability is variant of rotation angles. When rotation angle is coincident with object orientation, softmax probability has maximum value. During training, positive examples are rotated so that their orientation angles are zero. During detection, test images are rotated through different angles. At each angle, softmax probability is computed. A final object detection is based on maximum softmax probability as well as a pattern match between softmax probability patterns of all positive examples and the softmax probability pattern of a target object at different rotation angles. The object orientation is determined at the rotation angle when softmax probability has maximum value.