Object Detection and Segmentation for Abnormality Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing auto-encoder-based abnormality detection schemes struggle to differentiate between background noise and actual abnormalities, leading to reduced precision in detecting targets with positional and background variations.
Innovation Solution
An information processing apparatus that combines object detection and segmentation units, using neural networks to isolate targets and calculate index values that differentiate between normal and abnormal states, enhancing precision by normalizing pixel relationships within rectangular regions and pixel masks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If an auto-encoder is trained using only normal data to detect abnormality, then the scheme is highly versatile for detecting abnormality of targets, but it cannot differentiate between background noise and actual abnormalities, leading to reduced precision
Solution Approach 1:
The patent segments the input image into multiple regions including a target region and background regions. By processing these regions separately through the auto-encoder and comparing decoded values with original values independently, the system can distinguish between abnormalities in the target versus variations in the background, thereby resolving the contradiction between versatility and precision.
Solution Approach 2:
The patent applies different evaluation criteria to different regions of the image. The target region is evaluated for abnormality using one set of metrics, while background regions are evaluated separately to account for normal variations. This local differentiation allows the system to maintain high precision by not treating all regions uniformly, thus solving the contradiction.
2Adaptability or versatility
If the auto-encoder scheme is applied to targets with variations in position and background, then it can handle diverse scenarios, but the precision in detecting abnormality is lowered due to vulnerability to background and noise changes
Solution Approach 1:
By segmenting the image into target and background regions, the system can handle positional variations and background changes independently. The target region is extracted and processed separately, allowing the system to maintain precision even when the target position or background conditions vary across different input images.
Solution Approach 2:
The patent extracts the target region from the background using image processing techniques before feeding it to the auto-encoder. This extraction isolates the target of interest from distracting background elements, enabling the system to handle diverse scenarios with positional and background variations while maintaining high detection precision.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method includes: acquiring a training data set including pieces of training data, each of the pieces including an image of a training target, first annotation data representing a rectangular region in the image, and second annotation data; training, based on the image and the first annotation data, an object detection model specifying a rectangular region including the training target; training, based on the image and the second annotation data, a neural network; and calculating a first index value related to a relationship of a pixel number, the trained estimation model and the calculated first index value being used in a determination process that determines, based on the calculated first index value and a second index value relationship between a pixel number in an output result and an estimation result, whether or not a target in a target image is normal.