CNN Image Scaling Engine for Digital Magazine Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image scaling algorithms for digital magazines introduce errors and artifacts when increasing image resolution, leading to noise and haloing due to the need to guess new pixel values, especially when scaling to larger sizes.
Innovation Solution
A convolutional neural network (CNN) is used within an image scaling engine to preprocess and upscale images, utilizing convolutional and dense layers with linear Gaussian units, trained using supervised learning and backpropagation to minimize errors, without pooling operations to preserve feature location accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of moving object
If traditional interpolation algorithms (bicubic, bilinear, nearest-neighbor) are used to scale images to larger sizes, then the image can be enlarged, but errors, noise, haloing, and artifacts are introduced due to guessing new pixel values
Solution Approach 1:
The patent replaces traditional mechanical interpolation algorithms with a neural network-based system. The CNN learns to predict new pixel values through training on large datasets, substituting the deterministic mathematical interpolation processes with a data-driven probabilistic approach that better handles complex image structures and avoids artifacts.
Solution Approach 2:
The patent performs preliminary training of the CNN model using large datasets of image pairs (original and scaled versions) before actual image scaling. This preliminary learning phase allows the network to capture complex patterns and relationships in image data, enabling high-quality scaling without artifacts during actual operation.
2Ease of manufacture
If pixel values are kept in original color range (0-255), then the image data is simple to represent, but computational load increases during CNN operations due to large numbers
Solution Approach 1:
The patent transforms the parameter range of pixel values from the standard 0-255 range to a normalized range (typically -1 to 1 or 0 to 1). This parameter change reduces the magnitude of numbers involved in CNN computations, decreasing computational load and improving numerical stability during training and inference operations.
3Reliability
If pooling operations are used after convolutional layers, then classification invariance is improved, but feature location accuracy deteriorates which is critical for image scaling
Solution Approach 1:
The patent removes (extracts) the pooling operations from the CNN architecture. By eliminating the pooling layers that provide classification invariance, the network preserves exact spatial relationships and feature locations, which are critical for accurate image scaling where maintaining the precise position of image features is essential.
Data Source
AI summary
An online content system, such as a digital magazine, includes an image scaling engine for increasing the resolution of images. The image scaling engine comprises a convolutional neural network. An input image is preprocessed for use as inputs to a convolutional neural network (CNN). The preprocessed input image pixel values are used as inputs to the CNN. The CNN comprises convolutional layers and dense layers for determining image features and increasing image resolution. The CNN is trained using backpropagation to adjust model weights and biases. Each convolutional layer of a CNN detects features in an image by comparing image subregions to a set of known kernels and determining similarities between subregions and kernels using a convolution operation. The dense layers of the CNN have full connections to all of the outputs of a previous layer to determine the specific target output result such as output image pixel values.