Neural Network Image Retrieval System Using Output Value Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image retrieval systems face challenges in achieving high-accuracy and fast image retrieval, particularly in handling shape changes in images and efficiently processing large datasets.
Innovation Solution
An image retrieval system utilizing a neural network with a convolutional layer and a pooling layer, which compares output values from these layers for query and database image data to extract and rank images with high correspondence, enabling efficient and accurate retrieval.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature-based matching is used to extract feature points from images, then image retrieval accuracy can be improved, but processing time increases
Solution Approach 1:
The patent replaces traditional mechanical feature extraction methods with a neural network-based system. The neural network automatically learns and extracts meaningful features from images, eliminating the need for manual feature point extraction algorithms. This substitution enables both high accuracy and fast processing by leveraging the neural network's pre-trained knowledge of image patterns and structures.
Solution Approach 2:
The patent changes the fundamental parameters of image processing by using neural network output values instead of traditional feature point coordinates and descriptors. The neural network transforms images into a different representation space where similarity can be efficiently computed, fundamentally changing how image matching is performed and achieving both speed and accuracy improvements.
2Productivity
If area-based matching is used to compare template images with query images, then processing speed can be improved, but accuracy deteriorates especially when image shapes change
Solution Approach 1:
The patent changes the parameter space in which image comparison occurs by using neural network output values. Instead of comparing pixel values or feature coordinates directly, the system compares transformed representations from the neural network, which are invariant to shape changes. This parameter transformation enables both fast processing and high accuracy even when image shapes vary.
Solution Approach 2:
The neural network acts as an intermediary between the query image and database images. It processes both images through the same network, producing output values that capture essential visual characteristics. This intermediary transformation enables meaningful comparisons without direct pixel-wise or feature-point matching, solving the shape invariance problem while maintaining speed.
3Device complexity
If traditional template matching methods are used to search through large image databases, then system simplicity is maintained, but retrieval accuracy and speed both deteriorate
Solution Approach 1:
The patent replaces simple template matching algorithms with a neural network-based system. The neural network automatically learns hierarchical representations of images, enabling the system to handle complexity internally while presenting a simple interface. This substitution achieves high accuracy and speed without requiring complex manual configuration, as the neural network learns optimal features automatically.
Solution Approach 2:
The neural network performs preliminary processing by pre-extracting and organizing visual features during the training phase. This preliminary action creates a structured representation space where images are pre-processed and organized, allowing for fast and accurate retrieval during the actual search operation without requiring complex real-time processing.
Data Source
AI summary
An image retrieval system that enables high-accuracy image retrieval in a short time is provided. The image retrieval system includes a processing portion provided with a neural network. The neural network includes a layer provided with a neuron. The processing portion has a function of comparing query image data with a plurality of pieces of database image data, and extracting the database image data including an area with a high degree of correspondence to the query image data as extracted image data. The processing portion has a function of extracting data of the area with a high degree of correspondence to the query image data from the extracted image data, as partial image data. The layer has a function of outputting an output value corresponding to the features of the image data input to the neural network. The processing portion has a function of comparing the above output values in the case where the respective pieces of partial image data are input with the above output value in the case where the query image data is input.