Image Recognition Model Preprocessing for Overfitting Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As network models deepen, overfitting becomes more severe, leading to reduced performance, and existing regularization methods that randomly discard features often include irrelevant background information, resulting in unstable regularization effects.
Innovation Solution
Determine a target object positioning box covering effective features in an image sample and adjust pixel values within this box to create a preprocessed image sample, training the network model on this sample to enhance feature concentration and prominence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If network model depth increases to improve recognition capability, then recognition accuracy is improved, but overfitting becomes more severe
Solution Approach 1:
The patent extracts and removes part of the feature information within the target object positioning box by adjusting pixel values to a second value (e.g., setting to 0 or 255). This extraction of specific feature portions creates a preprocessed sample that maintains the target object while removing some discriminative features, thereby reducing overfitting while preserving recognition capability.
Solution Approach 2:
The patent changes the pixel value parameter within the target object positioning box from a first value to a second value. This parameter transformation modifies the feature representation in a controlled manner, creating data augmentation that prevents overfitting while maintaining the structural integrity of the target object for accurate recognition.
2Reliability
If existing regularization method randomly discards features to reduce overfitting, then overfitting is reduced, but irrelevant background information is included resulting in unstable regularization effects
Solution Approach 1:
The patent applies different processing to different regions of the image: within the target object positioning box, pixel values are adjusted to remove features, while regions outside the box maintain their original values. This localized processing ensures that background information outside the target area is preserved without introducing noise, while features within the target area are selectively removed to prevent overfitting, achieving stable regularization effects.
Solution Approach 2:
The patent performs preliminary determination of the target object positioning box before applying the regularization processing. By first identifying the precise location and boundaries of the target object, the method ensures that subsequent feature removal operations are accurately targeted, preventing both overfitting within the target area and unwanted modification of background regions, thus stabilizing the regularization effect.
Data Source
Figure 1~2
Figure 3~4
Figure 5~7
AI summary
Embodiments of the present application provide a processing method for an image recognition model and a related product. The processing method for the image recognition model includes: obtaining an image sample; determining a target object positioning box of the image sample, where the target object positioning box covers an area where an effective feature of the image sample is located; adjusting a pixel value of at least one pixel within the target object positioning box to obtain a preprocessed image sample; and training a network model based on the preprocessed image sample to obtain the image recognition model. The present application improves the concentration of the effective feature in the sample image and thus improves the regularization effect and reduces overfitting of the network model.