CNN Subnet Routing for Object Identification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for identifying specific objects in digital images, such as those depicting humans or animals, are computationally intensive and often limited to specific classes of objects, making them inefficient for surveillance systems and other applications that require identification across multiple classes.
Innovation Solution
A method using a convolutional neural network with a classification subnet and feature vector generating subnets to determine the class and generate feature vectors for objects, allowing for efficient comparison with registered vectors in a database to locate candidate images, which can handle multiple classes of objects and reduce computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithms using convolutional neural networks are used to identify objects, then identification accuracy is improved, but computational intensity increases significantly
Solution Approach 1:
The CNN is divided into multiple subnets, each responsible for identifying specific object classes. This segmentation allows the system to process images more efficiently by distributing computational tasks across specialized subnets rather than using a single monolithic network for all object types.
Solution Approach 2:
The system pre-trains multiple subnets on different object classes beforehand. When a new image is processed, the classification subnet quickly determines which pre-trained subnet to use, avoiding the need to train a new model from scratch and significantly reducing computational intensity for real-time identification.
2Measurement precision
If a single CNN is trained for a particular class of objects, then identification accuracy for that class is improved, but the system cannot identify objects from different classes
Solution Approach 1:
The system employs a universal CNN architecture that can identify multiple object classes by incorporating multiple subnets, each specialized for a particular class. The classification subnet acts as a router that directs images to the appropriate specialized subnet, enabling the single system to handle diverse object types while maintaining high accuracy for each class.
Solution Approach 2:
The system dynamically selects which subnet to use based on the classification result. Rather than statically assigning a single function to the entire CNN, the system adapts its computational path in real-time based on what type of object is being detected, optimizing both accuracy and efficiency for the specific task at hand.
3Adaptability or versatility
If manual methods are used to identify objects in digital images, then flexibility for different object types is maintained, but time consumption and labor requirements increase
Solution Approach 1:
Instead of manually analyzing each image, the system creates and uses digital copies (feature vectors) of objects for comparison. The CNN extracts key features and creates vector representations that can be quickly compared against stored templates, automating the identification process while maintaining the flexibility to handle various object types that would traditionally require manual analysis.
Data Source
AI summary
A method for finding one or more candidate digital images being likely candidates for depicting a specific object comprising: receiving an object digital image depicting the specific object; determining, using a classification subnet of a convolutional neural network, a class for the specific object depicted in the object digital image; selecting, based on the determined class for the specific object depicted in the object digital image, a feature vector generating subnet from a plurality of feature vector generating subnets; determining, by the selected feature vector generating subnet, a feature vector of the specific object depicted in the object digital image; locating one or more candidate digital images being likely candidates for depicting the specific object depicted in the object digital image by comparing the determined feature vector and feature vectors registered in a database, wherein each registered feature vector is associated with a digital image.