Object Detection Accuracy via Resolution-Specific Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep convolutional neural network algorithms for image recognition struggle with accuracy when dealing with low-resolution images due to their training on high-resolution images, and training on both resolutions leads to complex, non-convergent models.
Innovation Solution
The method involves separately training recognizing engines for high and low resolution images, using a categorizing engine to determine the image mode based on parameters like gain value or exposure time, and employing appropriate engines to process filtered images for accurate object detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep convolutional neural network models are trained on high resolution images, then image recognition accuracy is improved, but the model fails to accurately recognize low resolution testing images
Solution Approach 1:
The patent segments the image recognition task by creating separate deep convolutional neural network models for high resolution images and low resolution images. Each model is independently trained on its corresponding resolution type, allowing specialized feature extraction optimized for each resolution category. The system then segments the input image stream by resolution and routes to the appropriate model.
Solution Approach 2:
The patent changes the training parameters of the neural network models by using different input resolutions. High resolution models are trained with larger image dimensions and corresponding feature extraction parameters, while low resolution models use smaller dimensions and adjusted pooling/convolution parameters. This parameter differentiation enables each model to optimize for its specific resolution range.
2Measurement precision
If separate recognizing engines are trained for high and low resolution images, then recognition accuracy for both resolutions is improved, but device complexity increases
Solution Approach 1:
The patent creates a universal object detection system that handles both high and low resolution images through a unified architecture. The system uses a single object detection device that receives images of varying resolutions and automatically routes them to the appropriate pre-trained model. This multi-functional design allows one system to serve multiple resolution requirements without requiring separate complete detection pipelines.
Solution Approach 2:
The patent creates simplified copies of the deep learning models for low resolution images. Instead of using the full high resolution model architecture on downsampled images, it trains compact low resolution models that replicate the essential feature extraction capabilities at reduced computational complexity. These copied models are optimized for their specific resolution and can be deployed with fewer resources.
3Adaptability or versatility
If deep learning models are trained on both high and low resolution images, then both resolutions can be processed, but the model becomes complex and fails to converge
Solution Approach 1:
The patent segments the training process into separate batches for high resolution and low resolution images. Each segment is trained independently with resolution-appropriate hyperparameters, learning rates, and data augmentation strategies. This prevents the conflicting gradient signals from mixed resolution training from causing divergence, while still achieving versatility through multiple specialized models.
Solution Approach 2:
The patent implements dynamic model selection at inference time based on the input image resolution. The system dynamically chooses which pre-trained model to use (high resolution model or low resolution model) based on the detected input resolution, rather than using a single static model. This dynamic approach maintains convergence stability while achieving resolution flexibility.
Data Source
AI summary
The present disclosure discloses an object detection method based on machine learning. The object detection method senses a motion of an object by a motion sensor to generate a testing image, captures the testing image by an image sensor to transmit a sensed image to an object detection device, and detects a position of the object in the sensed image by the object detection device. Therefore, the present disclosure increases the accuracy of image recognition under various conditions.