Neural Network Loss Weighting for False Area Detection Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional detection techniques using neural networks often result in erroneous detections of areas that are not the target, particularly due to similarities in features between the target and surrounding regions, leading to inefficiencies in suppressing such errors.
Innovation Solution
A neural network training method that applies different weights to differences within and outside a designated area, with higher weights for errors within the area including the detection target, to enhance suppression of erroneous detections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional detection techniques are used, then the detection process is simple, but erroneous detections occur frequently
Solution Approach 1:
The patent applies local quality by differentiating between weighting areas (where the detection target is located) and non-weighting areas. The loss function assigns different weights to errors in different regions: a second weight is applied to differences within the weighting area, while a first weight is applied to differences outside the weighting area. This local differentiation enables the system to focus training attention on suppressing erroneous detections in critical regions without uniformly increasing complexity across the entire image.
2Reliability
If all areas are treated equally in training, then the training process is simple, but erroneous detections in specific areas cannot be suppressed efficiently
Solution Approach 1:
The patent implements local quality through the weighting scheme that distinguishes between weighting areas and non-weighting areas. The loss function calculates different loss values based on the spatial location of errors: errors within the weighting area (where the detection target is expected to be) are penalized more heavily with a second weight, while errors outside this area use a first weight. This localized approach enables efficient suppression of erroneous detections in critical regions.
Solution Approach 2:
The patent applies segmentation by dividing the image into a weighting area (containing the detection target) and a non-weighting area (excluding the detection target). This spatial segmentation allows the training system to treat different regions differently, focusing computational resources on suppressing erroneous detections within the weighting area while maintaining simpler processing for other regions.
Data Source
AI summary
An information processing apparatus includes at least one processor and at least one memory that is in communication with the at least one processor. The at least one memory stores instructions for causing the at least one processor and the at least one memory to train a neural network to detect target areas in images using training data, acquire object areas containing the detection target from the training data, set a weighting area based on these object areas, calculate a loss value based on the difference between the neural network's detection results and the training data. A second weight is applied to differences within the weighting area to calculate the loss value, causing the loss value to be larger than when using a first weight applied to differences outside this area, and the neural network is trained based on this loss value.