Fingerprint Image Reconstruction Using CNN Encoding-Decoding

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing fingerprint collection under optical screens face challenges in obtaining clear images due to overlapping and scattered fingerprint images from densely arranged pointolites, leading to inefficient and lengthy image collection processes.

Innovation Solution

A model training method using a convolutional neural network with encoding and decoding hierarchies, including asymmetrical convolution kernels and skip connections, to process blurred fingerprint images, enhancing image extraction accuracy and efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If densely arranged pointolites are used to collect fingerprint images, then the collection area and productivity are improved, but the obtained images become overlapping and scattered, reducing image quality

Engineering Contradiction:

Improvefingerprint collection efficiencyVSAvoidfingerprint image clarity

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent converts the harmful overlapping and scattered fingerprint images into a beneficial training dataset for the neural network. The model learns to separate and reconstruct clear fingerprint images from these degraded inputs, turning the collection efficiency problem into an opportunity for advanced image processing.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The neural network model acts as an intermediary between the blurred input images and the desired clear fingerprint images. It processes the degraded images through multiple encoding and decoding hierarchies to produce high-quality output, mediating the conflict between collection speed and image quality.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If traditional image processing methods are used to clear blurred fingerprint images, then the process is simple, but the accuracy and detail preservation are insufficient

Engineering Contradiction:

Improveimage extraction accuracyVSAvoidprocessing system complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces traditional mechanical/image processing methods with a neural network-based system. Instead of using conventional filtering or enhancement techniques, the system uses learned patterns from training data to reconstruct images, achieving superior detail preservation and accuracy.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The neural network dynamically adjusts processing parameters during inference based on the input image characteristics. The model learns optimal processing strategies from diverse training samples and adapts them to each specific input, enabling high accuracy across varying image conditions.

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If multiple encoding and decoding hierarchies are used in the neural network, then the image detail preservation is improved, but the computational complexity increases

Engineering Contradiction:

Improveimage detail preservationVSAvoidcomputational resource consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The neural network is segmented into multiple encoding hierarchies that extract features at different levels of abstraction, and multiple decoding hierarchies that reconstruct the image progressively. This segmentation allows the model to preserve details at various scales while distributing computational load across modular components.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The encoding and decoding hierarchies are nested within each other, with each hierarchy containing smaller processing units. This nested structure allows efficient feature reuse and progressive refinement, reducing redundant computations while maintaining high detail preservation capability.

Inventive Principle:

Principle #7Nested doll (Nesting)

Data Source

PatentUS12541825B2Model training method, image processing method, computing and processing device and non-transient computer-readable medium

Publication Date: 2026.02.03 BOE TECHNOLOGY GROUP CO LTD

AI summary

A model training method includes: acquiring a sample set, wherein samples in the sample set include a blurred image and a sharp image of a same fingerprint; inputting the blurred image into a convolutional neural network, performing, by an encoding network in the convolutional neural network, down-sampling and feature extraction to the blurred image, to output a plurality of feature maps, and performing, by a decoding network in the convolutional neural network, up-sampling and feature extraction to the feature maps, to output a predicted image corresponding to the blurred image; according to the predicted image, the sharp image and a predetermined loss function, calculating a loss value of the convolutional neural network, and, with minimizing the loss value as a target, adjusting parameters of the convolutional neural network; and determining the convolutional neural network of which the parameters are adjusted to be an image processing model.