Confidence Score Normalization for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image detection services using machine learning models often fail to accurately identify objects or misclassify images due to inconsistencies in confidence scores resulting from image manipulation techniques like scaling, flipping, and other transformations, which are not adequately accounted for in filtering methods.
Innovation Solution
The implementation of confidence score normalization and non-max suppression techniques to ensure accurate object detection across various image manipulations, where confidence scores are adjusted based on unmanipulated image distributions and overlapping bounding boxes are filtered to select the highest confidence scores, enhancing the model's ability to detect objects in manipulated images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If image manipulation techniques (scaling, flipping, transformation) are applied to improve object detection coverage, then detection coverage is improved, but confidence score consistency deteriorates
Solution Approach 1:
The patent applies parameter changes by transforming images through scaling, flipping, and other geometric operations to create augmented training data. These parameter transformations enable the model to learn from varied image configurations, improving detection coverage across different object orientations and scales while the normalization technique compensates for the resulting confidence score variations.
2Ease of manufacture
If models trained on non-augmented image sets are used with manipulated target images, then training simplicity is maintained, but detection accuracy deteriorates
Solution Approach 1:
The patent introduces confidence score normalization as an intermediary processing step between image manipulation and detection. This mediator technique adjusts confidence scores to account for transformation effects, enabling models trained on simple non-augmented datasets to maintain high detection accuracy when applied to manipulated target images without requiring complex retraining.
3Reliability
If multiple augmented versions of images are used for training, then model robustness is improved, but processing resources increase
Solution Approach 1:
The patent applies partial action by using confidence score normalization selectively on manipulated target images during inference rather than generating multiple augmented versions for every prediction. This approach achieves robust detection on transformed images without the excessive computational cost of processing multiple augmented variants, balancing model robustness with processing efficiency.
Data Source
AI summary
Computing systems may support image classification and image detection services, and these services may utilize object detection/image classification machine learning models. The described techniques provide for normalization of confidence scores corresponding to manipulated target images and for non-max suppression within the range of confidence scores for manipulated images. In one example, the techniques provide for generating different scales of a test image, and the system performs normalization of confidence scores corresponding to each scaled image and non-max suppression per scaled image These techniques may be used to provide more accurate image detection (e.g., object detection and/or image classification) and may be used with models that are not trained on modified image sets. The model may be trained on a standard (e.g. non-manipulated) image set but used with manipulated target images and the described techniques to provide accurate object detection.