Scene Recognition Model Training for Context-Aware Image Moderation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine moderation systems inaccurately determine the offensiveness of images due to ignoring scene context, particularly in cases where visual commonalities are present across different environments, such as anime or game scenes.
Innovation Solution
A scene recognition model is trained using a core feature extraction layer, global information feature extraction layer, local supervised learning (LCS) modules with attention mechanisms, and a fully-connected decision layer, employing multi-level training strategies to enhance feature extraction and scene recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine moderation relies on visual commonality of images, then processing efficiency is improved, but scene recognition accuracy deteriorates
Solution Approach 1:
The patent segments the image processing task into two independent parts: machine moderation handles the efficient classification based on visual commonality, while scene recognition separately analyzes spatial relationships and contextual information. This segmentation allows each component to optimize for its specific function without compromising the other, resolving the contradiction between processing efficiency and scene recognition accuracy.
Solution Approach 2:
The patent introduces scene recognition as an intermediary module that bridges machine moderation and final judgment. This intermediary analyzes the spatial relationships and contextual information between objects, providing nuanced insights that modify the machine moderation results without replacing the efficient automated classification process, thus maintaining both efficiency and accuracy.
2Extent of automation
If machine moderation uses visual commonality judgment, then automation level is improved, but judgment accuracy deteriorates
Solution Approach 1:
The patent segments the judgment process into automated machine moderation for initial classification and automated scene recognition for contextual analysis. Both segments operate automatically, maintaining high automation levels while improving reliability through the additional automated scene context analysis that prevents misclassification of objects in anime/game scenes.
Solution Approach 2:
The patent implements feedback by using scene recognition results to adjust and refine machine moderation judgments. The scene recognition module analyzes spatial relationships and provides feedback information that corrects inaccurate machine moderation results, such as identifying guns in anime/game contexts as non-offensive, thereby improving judgment accuracy while maintaining automation.
3Measurement precision
If scene recognition analyzes spatial relationships, then recognition precision is improved, but computational complexity increases
Solution Approach 1:
The patent segments the computational task into dedicated scene recognition modules that specifically process spatial relationships, separating this complex computation from the simpler machine moderation classification. This segmentation allows the system to manage computational complexity by processing different types of analysis in specialized modules rather than attempting to handle all processing requirements uniformly.
Solution Approach 2:
The patent applies partial action by focusing scene recognition computations only on the specific spatial relationships and contextual features necessary for accurate scene understanding, rather than performing exhaustive analysis of all possible image attributes. This selective computational approach improves recognition precision while controlling overall computational complexity by avoiding unnecessary processing.
Data Source
AI summary
Provided is a method for training a scene recognition model. The scene recognition model includes a core feature extraction layer, a global information feature extraction layer, an LCS module of at least one level with an attention mechanism, and a fully-connected decision layer. The method includes: acquiring parameters of the core feature extraction layer and the global information feature extraction layer by training based on a first scene label of a sample image and a standard cross-entropy loss; training a weight parameter of the LCS module of each level, based on a loss value acquired by performing a pixel-by-pixel calculation on a feature map output from the LCS module of each level and the first scene label of the sample image; and acquiring a parameter of the fully-connected decision layer by training based on the first scene label of the sample image and the standard cross-entropy loss.