Edge Pixel Compensation for Convolutional Neural Network Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolution neural networks struggle to extract edge features from images due to convolution kernels not extending beyond image edges, leading to loss of important information during information detection.
Innovation Solution
An image processing method that acquires a target image, counts its pixels to obtain a count value, and adds compensation data at the peripheral positions of edge pixels, creating a compensated target image that can be processed further for feature extraction and information detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If convolution neural network is used for information detection on image, then detection capability is improved, but edge features cannot be extracted due to convolution kernel not extending beyond image edge
Solution Approach 1:
The patent applies preliminary action by padding compensation data to the edge pixels of the input image before feeding it into the convolution neural network. This preliminary modification ensures that when the convolution kernel processes edge regions, it has access to sufficient contextual information from surrounding areas, thereby preventing edge feature loss while maintaining detection capability.
Solution Approach 2:
The patent changes the parameter of pixel data by introducing compensation data at edge positions. The compensation data is calculated based on the count value (number of pixels) and added to the edge pixel values, effectively modifying the input parameters to the convolutional layer to preserve edge information during the detection process.
2Ease of manufacture
If convolution kernel does not extend beyond image edge, then processing simplicity is maintained, but important information is lost during detection
Solution Approach 1:
The patent performs a preliminary padding operation before the convolution process. By pre-calculating and adding compensation data to edge pixels based on pixel count, the system maintains the simplicity of the convolution kernel structure while ensuring that edge regions contain sufficient information for accurate detection.
Solution Approach 2:
The compensation data acts as an intermediary element between the simple convolution kernel and the edge pixel regions. This intermediary padding layer provides the necessary contextual information to the convolution kernel without requiring the kernel itself to be complex or extend beyond the image boundaries.
Data Source
AI summary
An image processing method and device, an image detecting method and system, and a storage medium are provided. The image processing method comprises: acquiring a target image; counting pixels of the target image to obtain a count value; adding compensation data at a peripheral position of an edge pixel of the target image according to the count value, to obtain pixel data of the compensated target image.