Action Recognition Apparatus Using Addition Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing action recognition technologies, such as Two-Stream ConvNets, fail to accurately recognize actions of a target object when images other than the target object image overlap or are adjacent to the target object image.
Innovation Solution
An action recognition apparatus and method that generates an addition image by adding a preset setting image to a target object image, performs action recognition using this addition image, calculates recognition and estimation losses, and updates learning parameters to improve accuracy in recognizing actions of the target object.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Two-Stream ConvNets are used for action recognition, then action recognition can be performed using images and optical flow, but recognition accuracy deteriorates when other objects overlap or are adjacent to the target object
Solution Approach 1:
The patent segments the image processing task into two distinct streams: a spatial stream that processes static image features and a temporal stream that processes optical flow features. This segmentation allows each stream to specialize in extracting specific types of features, improving overall action recognition accuracy by separately handling spatial and temporal information before integrating them for final recognition
Solution Approach 2:
The patent introduces an intermediary attention mechanism that identifies and emphasizes the target object region while suppressing interference from overlapping or adjacent objects. This attention module acts as a mediator between the raw input images and the feature extraction process, guiding the network to focus on relevant regions and reducing the harmful impact of background objects
2Measurement precision
If deep learning models are trained with more data and complexity, then recognition accuracy improves, but computational cost and training time increase
Solution Approach 1:
The patent performs preliminary action by pre-processing images to generate optical flow fields before feeding them into the temporal stream. Additionally, the network architecture pre-defines the dual-stream structure and feature extraction pathways, allowing efficient processing during training and inference without requiring complex real-time computations
Solution Approach 2:
The patent adds the temporal dimension by introducing optical flow analysis alongside the spatial dimension of static images. This dimensional expansion allows the model to capture motion information and temporal dynamics of actions, significantly improving recognition accuracy by utilizing both spatial and temporal features from video sequences
Data Source
AI summary
An action recognition apparatus 1 including: a generation unit 2 that generates an addition image by adding a setting image to a target object image; an action recognition and estimation unit 3 that performs action recognition regarding the target object, outputs action recognition information indicating the result of action recognition, estimates the setting image, and outputs estimation information indicating the result of estimation; an action recognition loss calculation unit 4 that calculates an action recognition loss by using reference action recognition information generated in advance, and the action recognition information; an estimation loss calculation unit 5 that calculates an estimation loss by using reference estimation information generated in advance, and the estimation information; and a learning parameter updating unit 6 that updates learning parameters by using the action recognition loss and the estimation loss.