Deep-Bit Image Processing Using Clipped M-Bit CNN Segments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods using convolutional neural networks (CNNs) with INT8 precision fail to maintain tone when dealing with images having a bit depth deeper than 8 bits, leading to loss of image quality due to reduced bit precision.
Innovation Solution
An image processing apparatus that generates multiple images with M-bit depth per pixel by clipping segments from different positions within N-bit images and processes these through neural networks to produce an N-bit depth image, using a combination of first and second generation units to minimize tone reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a CNN with M-bit precision (e.g., INT8) is used for image processing, then computational efficiency is improved, but tone quality deteriorates due to reduction of bit precision from N-bit (e.g., 14-bit or 16-bit) to M-bit
Solution Approach 1:
The N-bit pixel value is divided into multiple M-bit segments by clipping from different positions. For example, a 14-bit pixel value is segmented into multiple 8-bit segments by clipping from the most significant bit position, middle position, and least significant bit position. These segmented images are then processed separately through the CNN and combined to reconstruct the N-bit output, thereby maintaining tone quality while using M-bit precision processing.
Solution Approach 2:
Multiple M-bit images are nested within the N-bit image structure. The patent generates multiple M-bit depth images from a single N-bit depth image by clipping segments from different positions, processes each M-bit image through the neural network, and then combines them to form the final N-bit output image. This nested approach allows the system to use efficient M-bit CNN processing while preserving N-bit tone quality.
2Manufacturing precision
If multiple M-bit images are generated from N-bit image by clipping segments from different positions, then tone quality is maintained, but processing complexity increases
Solution Approach 1:
The processing complexity is managed by segmenting the N-bit image into multiple M-bit segments that can be processed independently through the same neural network. This segmentation approach allows parallel processing of multiple segments and simplifies the overall architecture compared to developing a separate N-bit neural network, as the same M-bit CNN can be reused for each segment.
Solution Approach 2:
Instead of creating a complex N-bit neural network, the patent uses copying the M-bit neural network multiple times (or reusing it sequentially) to process different segments of the N-bit image. This copying approach reduces the complexity of network design and training, as the same proven M-bit network architecture can be applied to each segment without requiring new N-bit network development.
3Manufacturing precision
If M-bit segments are clipped from mutually different positions in N-bit image, then tone reduction is suppressed, but the number of processing steps increases
Solution Approach 1:
The N-bit image is segmented into multiple M-bit segments that can be processed in parallel or pipelined fashion. By clipping segments from different positions (most significant bit, middle bit, least significant bit), the patent enables concurrent processing of multiple segments through the neural network, reducing the overall processing time compared to sequential single-segment processing.
Solution Approach 2:
The patent performs preliminary clipping of M-bit segments from different positions of the N-bit image before neural network processing. This preliminary segmentation prepares the data in a format that allows efficient parallel processing through the M-bit CNN, reducing the total processing time compared to converting the entire N-bit image to M-bit or using a complex N-bit network.
Data Source
AI summary
An image processing apparatus including a first generation unit configured to generate a plurality of images having an M-bit depth per pixel by inputting an input image having a bit depth of N bits, and clipping M-bit segments from mutually different positions within the N bits; a second generation unit including a neural network for an M-bit-depth image and configured to generate, by inputting the plurality of images generated by the first generation unit into the neural network, a plurality of output images corresponding to each of the plurality of images being input; and a third generation unit configured to generate an image having the N-bit depth from the plurality of output images generated by the second generation unit.