Semantic Segmentation for Dominant Scene Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer vision systems fail to accurately identify the dominant scene or subject in images due to the lack of large and labeled datasets, leading to incorrect classification, especially in portable devices where computational resources are limited.
Innovation Solution
The method employs semantic segmentation using a convolutional neural network with an atrous spatial pyramid pooling module, spatial and class weighting, and temporal filtering to compute area ratios and determine the dominant class, which includes assigning importance weights based on object classes and locations within the image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic segmentation with complex processing is used to identify dominant scene, then classification accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The patent applies segmentation by dividing the image into multiple spatial regions (e.g., foreground, background, middle ground) and processing each region separately. This allows the system to focus computational resources on identifying dominant classes in specific regions rather than processing the entire image uniformly, thereby improving classification accuracy while reducing overall computational burden on mobile devices.
Solution Approach 2:
The patent implements local quality by applying different processing strategies to different spatial regions of the image. For example, foreground regions may receive more intensive processing to accurately identify dominant subjects, while background regions use simpler processing. This region-specific approach optimizes the balance between classification precision and computational efficiency on resource-constrained mobile processors.
2Measurement precision
If spatial and class weighting is applied to compute area ratios, then dominant class identification accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent changes parameters by introducing spatial weighting factors and class-specific weighting factors that modify the area ratio calculations. Instead of simply counting pixel areas, the system applies weighted sums where each pixel's contribution is adjusted by its spatial position and its class importance. This parameter transformation improves dominant class identification by emphasizing relevant regions and classes while maintaining a relatively straightforward computational framework suitable for mobile devices.
3Stability of the object's composition
If temporal filtering is applied across image sequences, then prediction stability is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by performing temporal filtering using previously computed results from earlier frames in the sequence. Instead of processing each frame completely independently, the system leverages temporal correlations by smoothing predictions across frames, where the dominant class identification in the current frame is influenced by results from previous frames. This approach stabilizes predictions and reduces flickering or inconsistent classifications while maintaining real-time processing capability on mobile devices.
Data Source
AI summary
A method for computing a dominant class of a scene includes: receiving an input image of a scene; generating a segmentation map of the input image, the segmentation map being labeled with a plurality of corresponding classes of a plurality of classes; computing a plurality of area ratios based on the segmentation map, each of the area ratios corresponding to a different class of the plurality of classes of the segmentation map; and outputting a detected dominant class of the scene based on a plurality of ranked labels based on the area ratios.