Convolution Operation Circuit with Rotation-Adaptive Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional convolutional neural networks are not robust to rotated images, requiring additional computational resources or hardware for image preprocessing or expanded training datasets, which increases costs and model size.
Innovation Solution
A convolution operation circuit that adjusts the convolution kernel based on the rotation state of input images, allowing direct convolution operations without additional preprocessing or retraining, using a rotation determination unit, convolution kernel adjustment unit, and convolution operation unit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image preprocessing (affine transformation) is performed to correct rotation, then the model's recognition accuracy is improved, but additional computational resources and hardware costs are required
Solution Approach 1:
The patent extracts the rotation correction function from the traditional image preprocessing stage and transfers it to the convolution operation stage. Instead of performing affine transformation on the entire image before input, the system applies rotation correction selectively during convolution operations by adjusting convolution kernels based on detected rotation angles, thereby eliminating the need for dedicated preprocessing hardware.
Solution Approach 2:
The patent introduces convolution kernel adjustment as an intermediary mechanism between image input and feature extraction. Rather than directly preprocessing images or using complex hardware, the system mediates the rotation issue by dynamically adjusting convolution kernels to compensate for rotation, achieving correction through parameter modification rather than structural transformation.
2Adaptability or versatility
If training data sets are expanded with rotated images, then the model's generalization ability is improved, but the model size increases and computational resource costs during training and inference increase
Solution Approach 1:
The patent changes the parameters of existing convolution kernels based on detected rotation angles rather than training with numerous rotated image variants. By adjusting kernel parameters (weights and biases) according to rotation state, the system achieves adaptability to different orientations without expanding the model's parameter space or requiring retraining on augmented datasets.
Solution Approach 2:
The patent introduces dynamic adjustment of convolution kernels based on real-time rotation detection. Instead of a static model trained on fixed datasets, the system dynamically adapts kernel parameters during inference to match the rotation state of input images, enabling generalization without expanding training data or model size.
3Ease of operation
If additional dedicated circuits are set up for image preprocessing, then rotation correction is achieved, but computational resource costs and hardware costs increase
Solution Approach 1:
The patent merges the rotation correction function with the existing convolution operation unit. Instead of adding separate preprocessing circuits, the system combines rotation detection and correction capabilities into the convolution layer itself, allowing the same hardware resources to serve multiple functions and eliminating redundant computational paths.
Data Source
AI summary
A convolution operation circuit includes: a rotation determination unit, a convolution kernel adjustment unit and a convolution operation unit. The rotation determination unit is configured to determine a rotation state of input data. The convolution kernel adjustment unit is coupled to the rotation determination unit and configured to selectively adjust an initial convolution kernel according to the rotation state, thereby obtaining an adjusted convolution kernel. The convolution operation unit is coupled to the convolution kernel adjustment unit and configured to perform a convolution operation based on the adjusted convolution kernel and the input data to obtain a feature map.