Image Processing Network Training via Markov Chain Cropping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Gaze estimation performed by trained image processing networks has low accuracy due to redundant pixels in training images.
Innovation Solution
A method for training an image processing network that involves determining a reference pixel in a training image annotated with a truth value, modeling the cropping operation as a Markov chain to determine cropping probabilities, and adjusting network parameters based on output results to obtain a trained network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If channel pruning is performed on the network using Differentiable Markov Channel Pruning (DMCP), then model acceleration and compression are achieved, but gaze estimation accuracy deteriorates due to redundant pixels in training images
Solution Approach 1:
The patent segments the image processing task by dividing the image into multiple patches and processing each patch independently through the neural network. This segmentation allows the model to focus on local features while maintaining overall accuracy, resolving the contradiction between compression and accuracy by processing images in manageable units rather than as a whole
Solution Approach 2:
The patent applies local quality by using overlapping patches that capture local image regions with different contexts. Each patch is processed with the pruned network, and results are aggregated to maintain global accuracy. This local processing approach compensates for the information loss from channel pruning while maintaining productivity benefits
2Reliability
If the image processing network processes the entire training image, then complete information is available for processing, but processing time and computational resources increase
Solution Approach 1:
The image is divided into multiple overlapping patches that are processed independently and in parallel. This segmentation reduces the computational burden on the network while maintaining processing completeness through the overlapping regions that capture contextual information from different perspectives
Solution Approach 2:
The patent processes multiple overlapping patches beyond the minimal required coverage. This excessive action ensures that all regions of the image are captured from multiple contexts, maintaining reliability while the parallel processing of patches reduces overall processing time compared to processing the entire image sequentially
Data Source
AI summary
Provided are a method and device for training an image processing network, and an image processing method and device. The method for training an image processing network includes following. A reference pixel is determined based on a training image annotated with a truth value. With the reference pixel as a starting point and based on a Markov chain of the training image, cropping probabilities of the image processing network processing the training image are determined. A network parameter value and the cropping probabilities of the image processing network are adjusted based on an output result obtained by the image processing network processing a training cropped area and the truth value, to obtain a trained image processing network. The training cropped area is obtained by cropping the training image based on the cropping probabilities.