Video Frame Selection for Diverse Prototype-kNN Patch Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models are vulnerable to patch attacks that degrade accuracy by presenting adversarial patches, and it is difficult to obtain suitable samples for use in prototype-kNN due to issues like similar image features and increased processing time.
Innovation Solution
A method involving segmentation of frames to identify regions reflecting objects, calculating overlap and similarity indices, and excluding unsuitable frames to obtain diverse samples suitable for prototype-kNN, thereby improving detection accuracy and reducing processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple samples are prepared for prototype-kNN to detect patch attacks, then detection accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-processing video frames to identify and exclude unsuitable samples before they are used in prototype-kNN. The system performs initial segmentation, overlap calculation, and similarity assessment to pre-filter frames, ensuring that only high-quality samples are retained. This pre-processing step reduces the number of samples that need to be processed during detection, thereby maintaining detection accuracy while significantly reducing processing time.
Solution Approach 2:
The patent applies segmentation by dividing the video stream into individual frames and further segmenting each frame into multiple candidate regions. This fine-grained segmentation allows the system to evaluate each region independently for suitability as a prototype sample, enabling selective retention of high-quality regions while discarding unsuitable ones. This segmentation approach facilitates efficient sample selection without compromising detection accuracy.
2Quantity of substance
If frames with similar image features are selected as samples, then the number of samples increases, but detection performance degrades
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the similarity threshold based on the specific characteristics of each video sequence and attack type. Instead of using a fixed similarity criterion, the system adapts the threshold parameter to optimize the balance between sample diversity and detection performance. This allows the system to maintain a high number of samples while ensuring they remain diverse and informative for detection.
Solution Approach 2:
The patent applies feedback by using the results of similarity calculations to iteratively refine sample selection. The system calculates similarity between candidate frames and already-selected prototypes, then uses this feedback information to determine whether to include or exclude each candidate. This feedback mechanism ensures that the final sample set maintains high diversity and avoids redundancy, thereby preserving detection performance while maximizing the number of useful samples.
3Reliability
If strict criteria are applied for sample selection, then sample quality improves, but the number of available samples decreases
Solution Approach 1:
The patent applies dimensionality change by evaluating candidate samples across multiple dimensions simultaneously - including spatial overlap, temporal consistency, feature similarity, and diversity metrics. Rather than relying on a single strict criterion that would eliminate most candidates, the system uses multi-dimensional assessment to identify samples that satisfy comprehensive quality requirements. This approach maintains high sample quality while preserving a sufficient number of candidates for effective detection.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
For each pair of two chronologically contiguous frames 111, an information processing device 100 calculates a first index value representing a degree of overlap between regions each identified from a corresponding one of the paired two frames 111. For each pair of two sequentially contiguous frames 111, the information processing device 100 calculates a second index value representing a degree of similarity between the paired two frames 111 with respect to a predetermined type of feature. To identify remaining frames 111, the information processing device 100, excludes from a series of frames 111, a certain number of frames 111 anterior and posterior to a pair at least whose first index value is less than a first threshold value or whose second index value is less than a second threshold value. The information processing device 100 extracts a frame 111 whose degree of similarity in the predetermined type of feature to another frame among the remaining frames 111 satisfies a predetermined condition.