Tile-Based CNN for Handwriting Detection in Image Forming Apparatus
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems face challenges in accurately and efficiently detecting and processing handwritten parts from images, particularly in terms of resource management and threshold value settings for removal and emphasis operations.
Innovation Solution
The system employs a convolutional neural network with an encoder and decoder architecture, dividing images into tiles to reduce RAM usage and using specific threshold values for determining handwriting probability, allowing for efficient detection and processing of handwritten parts by distinguishing between handwritten and non-handwritten areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the image is processed as a whole by the convolutional neural network, then the detection accuracy is improved, but the RAM usage increases significantly
Solution Approach 1:
The image is divided into multiple tile images (e.g., 222a, 222b, 222c, 222d) that can be processed independently by the convolutional neural network. This segmentation allows the system to process smaller portions of the image at a time, reducing the RAM requirements while maintaining detection accuracy through subsequent synthesis of the tile results into a complete segmentation image.
2Measurement precision
If the convolutional neural network processes all tile images through the decoder, then the detection accuracy is improved, but the processing time increases
Solution Approach 1:
The system performs preliminary determination of blank sheets by analyzing only the encoder output before proceeding to full decoder processing. This preliminary action allows the system to quickly identify and exclude blank sheet tiles from the more computationally intensive decoder processing, reducing overall processing time while maintaining accuracy for non-blank tiles.
Solution Approach 2:
Blank sheet determination is extracted as a separate, simplified processing path that bypasses the full decoder. By extracting this specific function and handling it separately, the system avoids unnecessary decoder processing for blank sheets, thereby reducing processing time without compromising the detection accuracy for actual handwritten content.
3Device complexity
If a single threshold value is used for handwriting detection, then the system complexity is reduced, but the detection accuracy decreases
Solution Approach 1:
The system applies different threshold values for different operations: a first threshold value for removing handwritten parts and a second threshold value for emphasizing handwritten parts. This local quality approach allows optimization of detection accuracy for each specific operation, with the first threshold optimized for removal and the second threshold optimized for emphasis, rather than using a single compromise threshold.
Data Source
AI summary
A CNN of an image forming apparatus includes: an encoder which compresses, for each tile image obtained by dividing an image into specific size pieces, information of the tile image; a decoder which restores the information of the tile image compressed by the encoder; and a blank sheet determination portion which determines whether the tile image corresponds to a blank sheet image. A segmentation image generation portion uses, when the blank sheet determination portion determines the tile image as being the blank sheet image, the blank sheet image for an image of a part corresponding to the tile image in a segmentation image, and uses, when the blank sheet determination portion determines the tile image as not being the blank sheet image, an output image of the decoder for an image of a part corresponding to the tile image in the segmentation image.