Salient Target Detection with Prior Features for Low-Quality Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing salient target detection networks have low detection accuracy on low-quality visual media data due to their high dependence on high-quality clean images.
Innovation Solution
An optimization method for constructing a target detection network that involves obtaining high-quality and low-quality visual media data, extracting features from both using preset prior and target detection networks, constructing feature correlation, salient target position, and salient prediction losses, and optimizing the target detection network based on these losses to improve detection accuracy on low-quality data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing salient target detection networks are constructed based on high-quality visual media data, then detection accuracy on high-quality data is improved, but detection accuracy on low-quality data deteriorates
Solution Approach 1:
A prior network is introduced as an intermediary component that processes high-quality visual media data to generate prior features, which then guide the target detection network in processing low-quality data. This intermediary bridge allows the system to leverage high-quality data characteristics when detecting targets in low-quality conditions, resolving the contradiction between high-quality data optimization and low-quality data adaptability
Solution Approach 2:
The prior network performs preliminary processing on high-quality visual media data to extract and store prior features before the actual detection task. These pre-computed prior features are then utilized during low-quality data detection, allowing the system to prepare advantageous information in advance that can be applied across different data quality conditions
2Measurement precision
If salient target detection networks rely on high-quality clean images, then feature extraction accuracy is improved, but feature consistency across different data qualities deteriorates
Solution Approach 1:
A feature correlation loss function is constructed that computes the correlation between prior features from the prior network and target features from the target detection network. This feedback mechanism guides the optimization process to maintain feature consistency across different data qualities while preserving extraction accuracy, as the loss function continuously adjusts features to maintain their correlation
3Productivity
If target detection networks are optimized for high-quality data, then detection performance on high-quality data is improved, but detection performance on low-quality data deteriorates
Solution Approach 1:
The target detection network is designed with multi-functionality to handle both high-quality and low-quality visual media data. By incorporating the prior network component and optimizing with the feature correlation loss, the same detection network structure achieves high performance across different data quality conditions, eliminating the need for separate specialized networks
Data Source
AI summary
Disclosed is an optimization method for constructing a target detection network. The method includes: obtaining high-quality visual media data, low-quality visual media data corresponding to the high-quality visual media data, and a corresponding true label, extracting a first backbone network side output feature generated by a preset prior network for the high-quality visual media data, and extracting a second backbone network side output feature generated by a preset target detection network to be trained for the low-quality visual media data; constructing a feature correlation loss, a salient target position loss and a salient prediction loss based on the first backbone network side output feature, the second backbone network side output feature, and the true label; and optimizing the preset target detection network to be trained based on the feature correlation loss, the salient target position loss, and the salient prediction loss to obtain the target detection network.