Multi-stage Deep Neural Network for Non-ideal Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition technologies face challenges in non-ideal environments due to localization, personalization, and scalability issues, particularly in recognizing food items in home or institutional settings, where images are poorly lit, uncentered, and of varying quality, leading to difficulties in object detection and subcategorization for nutritional assessment.
Innovation Solution
A multi-stage deep neural network framework that includes an object detection stage, a categorization stage, and a subcategorization stage, where the first DNN detects objects, the second DNN categorizes them, and the third DNN performs fine-grained subcategorization using localized training data from user input and ambient environment data, such as lighting and humidity, to produce accurate subcategorized data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single-stage deep neural network is used for image recognition, then the device complexity is reduced, but the measurement precision and reliability of object detection in non-ideal environments deteriorates
Solution Approach 1:
The patent divides the image recognition task into multiple sequential stages: object detection stage, categorization stage, and subcategorization stage. Each stage is handled by a separate deep neural network that processes the output of the previous stage, allowing specialized optimization for each function while maintaining overall system accuracy in non-ideal environments.
Solution Approach 2:
The patent transitions from single-stage to multi-stage processing, adding a temporal dimension to the recognition pipeline. This staged approach allows the system to progressively refine recognition results through multiple processing passes, improving precision without requiring exponentially complex single-stage architectures.
2Reliability
If deep neural networks are trained on stock image data from ideal environments, then the training data quality is high, but the adaptability to non-ideal real-world environments deteriorates
Solution Approach 1:
The patent applies local quality by training different neural networks with different data characteristics appropriate to their specific functions. The object detection network is trained on diverse images including non-ideal conditions, while subsequent categorization and subcategorization networks are trained on progressively more specific data, allowing each stage to optimize for its particular recognition task in the target environment.
Solution Approach 2:
The patent performs preliminary training on stock image data to establish baseline recognition capabilities, then applies fine-tuning with environment-specific data. This two-phase training approach allows the system to first learn general features from high-quality diverse data, then adapt to specific non-ideal environmental conditions through targeted retraining on locally captured images.
3Productivity
If coarse-grained categorization is performed first, then the processing speed is improved, but the subcategorization accuracy for fine-grained recognition deteriorates
Solution Approach 1:
The patent segments the categorization task into two distinct sequential operations: coarse-grained categorization followed by fine-grained subcategorization. This segmentation allows the system to first quickly identify broad object categories, then apply more computationally intensive fine-grained analysis only when needed, optimizing both speed and accuracy.
Solution Approach 2:
The patent applies partial action by performing full subcategorization analysis only for objects that require fine-grained distinction. The multi-stage architecture allows the system to quickly filter through coarse categories and apply detailed subcategorization selectively, rather than performing exhaustive analysis on all images, thus maintaining productivity while improving subcategorization precision where necessary.
Data Source
AI summary
Provided are an apparatus and a method of multi-stage image recognition. For the multi-stage image recognition, categorized object data is received from a first deep neural network. A second deep neural network is trained on subcategory customization data that relates to a non-ideal environment when the second deep neural network produces invalid subcategorized object data from the categorized object data, and generates an image recognition result using the second deep neural network as trained.