Windowed Contextual Pooling for Neural Network Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fully convolutional neural networks (FCNs) have a limited receptive field, making it difficult to detect and classify larger objects in images, particularly in document images that contain elements spanning significant portions of the image, and existing solutions like LSTM networks are computationally intensive and inefficient for mobile platforms.
Innovation Solution
The implementation of windowed contextual pooling (WCP) in neural networks, which increases the receptive field by using horizontal and vertical pooling layers with larger kernel sizes, allowing for improved object detection and classification by segmenting input feature channels, applying different pooling processes, and merging them with the original feature channels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If fully convolutional neural networks (FCNs) are used for object detection, then the network can process images efficiently, but the receptive field is limited making it difficult to detect large objects
Solution Approach 1:
The patent segments the feature channels into multiple groups and applies windowed pooling to each group independently. This segmentation allows the network to process large receptive fields by dividing the computation into manageable windows, thereby detecting large objects while maintaining processing efficiency.
Solution Approach 2:
The patent introduces windowed pooling operations that operate in both horizontal and vertical dimensions simultaneously. By applying pooling layers with larger kernel sizes in multiple dimensions, the network expands its receptive field capability without proportionally increasing computational complexity, thus resolving the contradiction between efficiency and receptive field size.
2Area of stationary object
If LSTM networks are used to increase receptive field, then larger objects can be detected, but computational intensity and memory requirements increase
Solution Approach 1:
The patent replaces LSTM networks with a windowed pooling approach that uses standard convolutional pooling operations. This substitution eliminates the need for complex recurrent computations while achieving the same receptive field expansion, thereby significantly reducing computational intensity and memory requirements.
Solution Approach 2:
The patent changes the approach by using pooling layers with larger kernel sizes instead of LSTM units. By adjusting the kernel size parameters of pooling layers, the network achieves large receptive fields through simple parameter modification rather than complex architectural changes, reducing computational and memory demands.
3Measurement precision
If pooling layers with larger kernel sizes are used to increase receptive field, then object detection accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the input feature channels into multiple groups and applies windowed pooling to each group independently. This segmentation allows the network to achieve high detection accuracy through multiple small pooling operations rather than a single large complex operation, reducing overall computational complexity.
Solution Approach 2:
The patent merges the results of multiple windowed pooling operations across different groups and dimensions. By combining the outputs of simpler pooling operations rather than using a single complex operation, the network achieves high accuracy while maintaining lower computational complexity through the merging of simple operations.
Data Source
AI summary
Techniques are disclosed for neural network based windowed contextual pooling. A methodology implementing the techniques according to an embodiment includes segmenting input feature channels into first and second groups of feature channels. The method also includes applying a first windowed pooling process to the first group of feature channels to generate a first group of pooled feature channels and applying a second windowed pooling process to the second group of feature channels to generate a second group of pooled feature channels. The method further includes performing a weighted merging of the first group of pooled feature channels and the second group of pooled feature channels to generate merged pooled feature channels. The method further includes concatenating the merged pooled feature channels with the input feature channels to generate concatenated feature channels and applying a two-dimensional convolutional neural network to the concatenated feature channels to generate contextually pooled output feature channels.