Composite Image Feature Extraction With Crop-Isolated CNN Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object re-identification or classification using convolutional neural networks (CNNs) face inefficiencies when processing multiple image crops as a composite image, leading to information leakage between crops due to padding requirements and increased processing time.
Innovation Solution
A method involving the concatenation of images into a composite image, with convolutional filters selectively applied to specific regions using a grid pattern to nullify weights that cover data from non-currently convolved images, thereby maintaining data separation and enhancing parallel computation efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple image crops are concatenated into a composite image for parallel processing, then processing capacity per time unit increases, but information leakage occurs between crops during convolution
Solution Approach 1:
The patent divides the composite image into distinct grid regions, each corresponding to a separate input image. By mapping convolutional filter positions to grid coordinates, the system processes each region independently, preventing information leakage while maintaining parallel processing efficiency.
Solution Approach 2:
The patent applies different processing characteristics to different regions of the composite image. Specifically, convolutional filter weights are selectively activated or deactivated based on which grid region the filter currently covers, allowing each region to be processed with appropriate local properties without affecting other regions.
2Loss of information
If padding is applied to prevent information leakage between crops, then information separation is maintained, but processing time increases
Solution Approach 1:
The patent dynamically adjusts convolutional filter behavior based on position within the composite image. Instead of static padding, the system actively controls weight activation during convolution operations, switching between different processing modes depending on the current grid region being processed.
Solution Approach 2:
The patent changes the effective parameters of convolutional filters during processing by selectively activating or deactivating weights based on grid position. This dynamic parameter adjustment prevents information leakage without requiring additional padding operations that would increase processing time.
3Measurement precision
If the number of convolutional layers is increased for better feature extraction, then classification accuracy improves, but the quantity of padding required increases, reducing the number of crops that can be combined
Solution Approach 1:
The patent introduces a new dimension for control by mapping the 2D spatial position of convolutional filters to grid coordinates. This additional dimensional mapping allows independent control of processing behavior across multiple layers without requiring proportional increases in padding, enabling deeper networks to process more crops simultaneously.
Data Source
Figure 1
Figure 2
Figure 3a~3b
AI summary
A method for feature extraction of detected objects, comprising the steps of: receiving a plurality of images, each depicting an object detected by the object detecting application; concatenating the plurality of images into a composite image according to a grid pattern; feeding the composite image through a convolutional neural network (CNN) trained for feature extraction, wherein each convolutional layer of the CNN is configured to, while convolving input data to the convolutional layer using a convolutional filter: determine a currently convolved image of the plurality of images by determining a centre coordinate of a subset of the input data currently covered by the convolutional filter, and mapping the centre coordinate to the grid pattern; and selectively nullifying all weights of the convolutional filter that cover input data derived from any of the plurality of images not being the currently convolved image.