Dynamic Convolutional Kernel Attention for Positional Feature Extraction

The patent transforms static convolutional kernels into dynamic ones by introducing attention mechanisms that adaptively adjust kernel parameters based on input data characteristics. The attention weights are computed from the input feature map and used to modulate the convolutional kernels, enabling the model to dynamically adapt to different input patterns while maintaining a relatively simple overall structure.

The patent introduces attention weights as an intermediary component between the input feature map and the convolutional kernels. These attention weights serve as a mediator that carries positional and contextual information from the input to the kernels, allowing the kernels to adapt their behavior based on the specific input without requiring complete restructuring of the model.

The patent applies attention mechanisms locally at each convolutional layer rather than globally across the entire network. The attention weights are computed based on local feature map characteristics and applied to local kernels, which improves feature extraction accuracy for specific regions while avoiding the excessive computational cost of global attention mechanisms.

The patent implements a simplified version of attention mechanisms that computes only the necessary attention weights for dynamic kernel adjustment, rather than implementing full attention mechanisms. This partial action approach provides sufficient adaptability for improving feature extraction while keeping computational overhead manageable by avoiding redundant calculations.