Learning method and learning device for object detector for surveillance purposes based on a neural folding network capable of converting modes according to object dimensions, as well as testing device and testing method therewith
Patent Information
- Application Number
- DE602019064186
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-01-25
- Filing Date
- 2019-11-11
- Publication Date
- 2025-01-01
- Estimated Expiration
- 2039-11-11
AI Technical Summary
Conventional object detection methods using convolutional neural networks (CNNs) face performance deterioration due to limited information from feature maps pooled with a single scale, leading to increased operations and reduced detection accuracy.
Implementing pooling layers with different scales to pool areas corresponding to regions of interest (ROIs) on feature maps, generating multiple feature vectors per scale, and using fully connected layers to output class and regression information, which are then learned through backpropagation to reduce the number of operations and improve detection performance.
This approach reduces the number of operations required for object detection, enhances feature utilization, and improves detection accuracy by allowing each pooling layer to perform pooling operations at varying scales, resulting in improved computational efficiency and performance.