Saliency Map Generation via Resolution-Aligned Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting salient objects in images using convolutional neural networks face challenges due to the direct fusion of feature maps from different layers, which results in suboptimal saliency detection outcomes due to differences in feature map resolutions and quality.
Innovation Solution
The method involves performing convolution processing on multiple convolutional layers to obtain feature maps with varying resolutions, followed by superposition and splicing processes to create a saliency map, where feature maps with lower resolutions are upscaled to match higher resolution maps, and weights are trained based on differences between saliency maps and reference maps to enhance fusion and feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If feature maps from different convolutional layers are directly fused, then the detection process is simple, but the saliency detection result is not ideal due to great differences in feature map qualities
Solution Approach 1:
The patent applies preliminary action by performing resolution alignment and quality enhancement on feature maps before fusion. Specifically, it resamples feature maps from different layers to a unified resolution and applies quality enhancement processing to ensure all feature maps have consistent quality metrics before being fused, thereby resolving the quality mismatch problem while maintaining a relatively simple overall process
Solution Approach 2:
The patent changes key parameters of feature maps including resolution (through resampling), quality metrics (through quality enhancement processing), and feature representation (through quality-aware fusion with learnable parameters). These parameter transformations enable effective fusion of multi-scale features while improving detection accuracy
2Ease of manufacture
If feature maps with different resolutions are directly fused, then the processing is straightforward, but the sparsity and irregularity of the result increase
Solution Approach 1:
The patent performs resolution alignment as a preliminary step before fusion by resampling feature maps from different convolutional layers to a unified resolution. This ensures that all feature maps have consistent spatial dimensions and regular structures before being combined, preventing sparsity and irregularity in the final result
Solution Approach 2:
The patent changes the resolution parameter of feature maps through resampling operations, transforming them to a common resolution scale. This parameter standardization enables regular and consistent feature map composition during fusion, eliminating the irregularity caused by direct fusion of multi-resolution features
3Manufacturing precision
If shallow layer feature maps are used, then fine details are preserved, but the feature maps become too messy and noisy
Solution Approach 1:
The patent applies quality enhancement processing as a preliminary step to shallow layer feature maps before fusion. This processing reduces noise and messiness while preserving the fine details and high-frequency information characteristic of shallow features, thereby improving their reliability for subsequent fusion operations
Solution Approach 2:
The patent changes the quality parameters of shallow layer feature maps through quality enhancement processing, which adjusts noise levels, sharpness, and other quality metrics. This transformation maintains the detailed information while improving the overall reliability and usability of shallow features in the fusion process
4Stability of the object's composition
If deep layer feature maps are used, then regularity is improved, but the feature maps lack fine details and become too abstract
Solution Approach 1:
The patent performs resolution alignment on deep layer feature maps as a preliminary step, resampling them to a unified resolution that preserves their inherent regularity while enabling proper integration with shallow features. This ensures deep features maintain their structural advantages during fusion
Data Source
AI summary
A method and an apparatus for detecting a salient object in an image includes separately performing convolution processing corresponding to at least two convolutional layers on a to-be-processed image to obtain at least two first feature maps of the to-be-processed image, performing superposition processing on at least two first feature maps included in a superposition set in at least two sets to obtain at least two second feature maps of the to-be-processed image, the at least two sets are in a one-to-one correspondence with the at least two second feature maps, and a resolution of a first feature map included in the superposition set is lower than or equal to a resolution of a second feature map corresponding to the superposition set, and splicing the at least two second feature maps to obtain a saliency map.