Neural Network Activation Subtraction for Occluded Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image-based classification systems face challenges in accurately classifying objects partially occluded by foreign objects due to noise introduction and computational inefficiencies in existing occlusion handling methods, such as manual labeling and costly preliminary steps for occlusion detection.
Innovation Solution
The method involves using a convolutional neural network with an activation region subtractor that determines and subtracts neuron activations associated with foreign objects within the network, allowing for accurate classification in a single forward pass and reducing computational expense.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If preliminary detection and removal of foreign objects is performed before classification, then classification accuracy is improved, but computational cost and processing time increase significantly
Solution Approach 1:
The patent merges the foreign object detection and removal step with the classification process by performing both operations within the same neural network forward pass. The activation subtractor is integrated into the CNN architecture, allowing simultaneous classification and occlusion handling without requiring separate preliminary processing steps, thereby reducing computational overhead while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network to recognize both target objects and foreign objects during the training phase. The network learns to identify activation patterns associated with foreign objects in advance, enabling it to subtract these activations during inference without requiring real-time detection algorithms, thus reducing processing time and computational cost.
2Adaptability or versatility
If multiple types of foreign objects are handled through manual labeling or automatic augmentation in training data, then occlusion handling capability is improved, but dataset construction difficulty and scalability worsen
Solution Approach 1:
The patent implements a universal approach by training the neural network to handle multiple types of foreign objects simultaneously through a single classification model. The network learns general occlusion patterns and features that apply across different foreign object types, eliminating the need to construct separate datasets for each occlusion scenario. The activation subtractor operates universally on any foreign object detected, providing scalable occlusion handling without increasing dataset construction complexity.
Data Source
AI summary
A method for classifying an occluded object includes receiving, by one or more processing circuits, an image of the object that is partially occluded by a foreign object and classifying, by the one or more processing circuits, the object of the image into one of one or more classes of interest via an artificial neural network (ANN) by determining a plurality of neuron activations of neurons of the ANN for one or more foreign classes and the one or more classes of interest, subtracting one or more of the neuron activations of the one or more foreign classes from the neuron activations of the one or more classes of interest, wherein the foreign object belongs to one of the one or more foreign classes, and classifying the object of the image into the one of the one or more classes of interest based on the subtracting.