Deep Variant Coding for Mobile Visual Search Data Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional mobile visual search methods are inefficient in low bandwidth environments due to high data transmission requirements and are computationally expensive, making them unsuitable for devices with limited resources, especially in ecommerce applications.
Innovation Solution
The implementation of deep variant coding (DVC) for image representation, which uses a set of hash functions to generate a variable-length encoding that reduces data transmission and requires fewer computational resources, allowing for efficient indexing and searching on both mobile devices and servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If compressed image format is transmitted to search server, then image data can be searched, but large amount of data transmission is required leading to unfavorable user experience in low bandwidth environments
Solution Approach 1:
The patent extracts only the essential visual features from images using deep neural networks and encodes them into compact hash signatures. Instead of transmitting or processing entire compressed images, only the extracted feature signatures are retained and searched, dramatically reducing data volume while preserving search capability
Solution Approach 2:
The patent transforms image data from pixel-space to feature-space representation through deep neural network processing. This parameter transformation converts high-dimensional image data into low-dimensional hash signatures that maintain discriminative power for visual search while occupying minimal storage and transmission bandwidth
2Quantity of substance
If feature extraction is performed on mobile device, then data transmission is reduced, but computational expense is high making it unsuitable for devices with limited resources
Solution Approach 1:
The patent performs feature extraction and hash signature generation in advance during image ingestion or preprocessing phases. Mobile devices can leverage pre-computed deep features from neural networks that are trained offline, avoiding the need to perform computationally intensive feature extraction during runtime visual search operations
Solution Approach 2:
The patent uses pre-computed deep feature representations and hash signatures that can be stored locally on mobile devices. Instead of performing real-time feature extraction, the system copies and utilizes pre-processed feature data that has been prepared beforehand, significantly reducing computational burden on resource-constrained mobile devices
3Device complexity
If conventional fixed-length hashing is used, then implementation is simple, but data representation is inefficient with denser regions requiring same bits as sparser regions
Solution Approach 1:
The patent implements dynamic, variant-length hash coding where the number of bits allocated to represent different image regions varies based on their information density. Denser regions with more visual detail are assigned more hash bits for finer discrimination, while sparser regions use fewer bits, creating an adaptive encoding scheme that optimizes the balance between representation accuracy and data compactness
Data Source
AI summary
An approach for performing mobile visual search uses deep variant coding of images to reduce the amount of data transmitted from mobile devices to a search server and to provide more efficient indexing and searching on the search server. The amount of data used to represent an image varies depending upon the content of the image and is less than conventional fixed bit length hashing approaches. Denser regions of a feature space are represented by more encoding bits and sparser regions of the feature space are represented by fewer encoding bits, so that the overall number of encoding bits for an image feature is reduced. The approach generally involves determining a set of hash functions that provide deep hashing with more evenly-distributed hash buckets. One or more additional hash functions may be selectively generated for particular hash buckets that contain more than a specified number of images.