Cascaded Convolutional Neural Network for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural networks (CNNs) for object detection in images require extensive computational resources and complex training processes due to the need for multiple layers and varying down-sampling ratios, which increases complexity and reduces efficiency.
Innovation Solution
A CNN system with multiple object detectors, each associated with a specific down-sampling ratio and image window size, sharing common convolution layers, and employing a training process that aligns samples with feature reference locations and perturbs them to increase the number of training samples, allowing for more efficient detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple object detectors with varying down-sampling ratios are employed to detect objects at different scales, then detection accuracy across varying image scales is improved, but computational complexity and device complexity increase
Solution Approach 1:
The patent merges multiple object detectors into a single integrated system where detectors with the same down-sampling ratio share common convolution layers. This combining approach allows the system to maintain multiple down-sampling ratios for detecting objects at different scales while reducing redundant computations through shared layers, thereby improving detection accuracy without proportionally increasing computational complexity
Solution Approach 2:
The patent creates universal convolution layers that serve multiple functions across different object detectors. The shared convolution layers perform feature extraction for multiple detectors simultaneously, enabling a single set of layers to handle detection at various scales through different down-sampling ratios, thus reducing overall device complexity while maintaining multi-scale detection capability
2Measurement precision
If multiple object detectors with varying down-sampling ratios are employed to detect objects at different scales, then detection accuracy across varying image scales is improved, but the number of parameters and training complexity increase
Solution Approach 1:
The patent combines multiple detectors that share the same down-sampling ratio into a unified structure with common convolution layers. This merging reduces the total number of parameters by eliminating redundant weight matrices and bias terms that would otherwise be replicated across separate detectors, while still maintaining the ability to detect objects at different scales through the shared architecture
Solution Approach 2:
The patent discards redundant parameters by sharing convolution layers across detectors with the same down-sampling ratio, and recovers the necessary detection capability through the shared feature representations. The common layers learn general features that are recovered and reused by multiple detectors, reducing parameter count while maintaining detection accuracy across scales
3Measurement precision
If extensive training data with multiple scales is used to improve detection accuracy, then detection performance across varying image scales is improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary feature extraction through shared convolution layers that learn general features applicable across multiple scales. By pre-learning these universal features in the shared layers, the system reduces the amount of scale-specific training needed, thereby improving detection performance across varying image scales without proportionally increasing training time and computational resources
Data Source
AI summary
A convolutional neural network system for detecting at least one object in at least one image. The system includes a plurality of object detectors, corresponding to a predetermined image window size in the at least one image. Each object detector is associated with a respective down-sampling ratio with respect to the at least one image. Each object detector includes a respective convolutional neural network and an object classifier coupled with the convolutional neural network. The respective convolutional neural network includes a plurality of convolution layers. The object classifier classifies objects in the image according to the results from the convolutional neural network. Object detectors associated with the same respective down-sampling ratio define at least one group of object detectors. Object detectors in a group of object detectors being associated with common convolution layers.