Server-Side Neural Network Image Recognition to Reduce False Positives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies for identifying retail products in supermarkets suffer from accuracy issues, particularly item confusion and false positives, and are not commercially viable due to battery life and data usage concerns when used on mobile devices.
Innovation Solution
Implement a neural network architecture with parallel output sections of different sizes for continuous training and classification, combined with image augmentation techniques and false positive mitigation strategies, to enhance accuracy and reduce computational load on mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image recognition is performed on mobile devices, then recognition accuracy can be improved, but battery life is reduced
Solution Approach 1:
The image recognition system is segmented into two parts: a lightweight client application on the mobile device that captures images and a server application that performs the computationally intensive neural network processing. This segmentation allows the mobile device to maintain accuracy through server-based processing while preserving battery life by avoiding local computational burden.
Solution Approach 2:
A communication network acts as an intermediary between the mobile device and the server. The mobile device transmits captured images through the network to the server for processing, and receives recognition results back. This intermediary enables accurate recognition without requiring the mobile device to perform heavy computations locally.
2Measurement precision
If image recognition is performed on mobile devices, then recognition accuracy can be improved, but data usage increases
Solution Approach 1:
The computationally intensive neural network processing is extracted from the mobile device and placed on the server. Only the essential image data is transmitted to the server, and the heavy computational workload is performed remotely, reducing the data transfer requirements compared to running complex models locally.
Solution Approach 2:
Instead of copying the entire neural network model to the mobile device, the system uses a simplified client application that communicates with the server. The server maintains the full neural network model and processes images remotely, eliminating the need to transfer large model files while maintaining recognition accuracy.
3Reliability
If conventional barcodes are used for product identification, then identification certainty is maintained, but checkout process time increases
Solution Approach 1:
The mechanical process of manually locating and presenting barcodes to scanners is replaced with an automated image recognition system. The mobile device camera captures the product packaging, and the neural network automatically identifies the product, eliminating the need for manual barcode manipulation and enabling faster checkout.
Solution Approach 2:
The system changes the input parameter from barcode symbols to product packaging images. This parameter change allows the system to recognize products based on their visual appearance rather than requiring specific barcode markings, thereby maintaining identification certainty while improving checkout speed.
4Productivity
If digital watermarks are used instead of barcodes, then checkout efficiency is improved, but production cycle time increases
Solution Approach 1:
The system performs preliminary actions by capturing images of product packaging before the checkout process. The neural network processes these images to identify products, enabling the checkout system to operate independently of the packaging production cycle. This allows retailers to transition to digital watermarking without being constrained by the production cycle timing.
Data Source
AI summary
Aspects of the detailed technologies concern training and use of neural networks for fine-grained classification of large numbers of items, e.g., as may be encountered in a supermarket. Mitigating false positive errors is an exemplary area of emphasis. Novel network topologies are also detailed—some employing recognition technologies in addition to neural networks. A great number of other features and arrangements are also detailed.