Object Detection Retraining Using Natural Misdetection Image Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing relearning methods for object detection devices do not consistently improve recognition performance due to the use of unnatural images generated through geometric transformations, which can adversely affect training effectiveness.
Innovation Solution
An information processing method that automatically generates training data by extracting misdetected object areas from images and superimposing or substituting them into similar object classes, ensuring a natural appearance and improving detection performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If geometric transformation is used to generate training image data, then the quantity of training data is increased, but the naturalness of the generated images deteriorates
Solution Approach 1:
The image is segmented into a background portion and an object portion. The background portion is retained from the original image to maintain naturalness, while the object portion is transformed and recombined. This segmentation allows the generation of diverse training data without compromising the natural appearance of the background environment.
Solution Approach 2:
The object portion is extracted from the original image through segmentation. This extracted object can then be transformed (rotated, scaled, etc.) and recombined with different backgrounds or in different configurations, enabling the generation of multiple varied training images from a single original image while preserving naturalness.
2Adaptability or versatility
If random shifting or rotation is applied during geometric transformation, then the diversity of training data is improved, but the training effect deteriorates due to unnatural appearance
Solution Approach 1:
Different parts of the image are treated differently: the background portion maintains its original natural appearance while the object portion undergoes transformation. This local differentiation allows the object to be diversified for better adaptability while the natural background preserves the training effect.
Solution Approach 2:
The image is pre-processed through segmentation before transformation. By separating the object from the background in advance, the subsequent geometric transformations can be applied selectively to the object portion only, ensuring that the background remains natural and the overall image maintains training effectiveness.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing method includes the following executed by a computer: acquiring a first image and object data of an object appearing in the first image (S10), extracting a portion of the first image that corresponds to a difference between the object data and an object detection result obtained by inputting the first image to a trained model (S60), the trained model receiving an image as input to output an object detection result, acquiring a second image that includes a portion corresponding to the same object data as object data corresponding to the extracted portion of the first image (S70), reflecting an image based on the extracted portion of the first image in the portion of the acquired second image that corresponds to the same object data (S80), and generating training data for the trained model (S90).