Entropy Estimation Network With Local Attention for Image Compression

The patent uses segmentation through group convolution where the feature map is divided into G independent groups. Each group is processed separately, allowing the network to capture broader contextual information (expanded receptive field) while maintaining manageable computational complexity. The segmentation enables parallel processing of different feature groups, effectively expanding the receptive field without linearly increasing computational burden.

The patent applies dimensionality change by using pointwise convolution with 1x1 kernels that operate in the channel dimension rather than spatial dimensions. This allows the network to process and integrate information across channels effectively, expanding the functional receptive field in the channel dimension while keeping spatial computational complexity controlled. This dimensional approach enables capturing more contextual information without proportionally increasing computational load.

The patent applies segmentation by implementing a two-stage convolution process where group convolution extracts features in a computationally efficient manner, followed by pointwise convolution that refines features with minimal computational overhead. This segmented approach allows the network to achieve deep feature extraction capabilities without proportionally increasing overall network complexity, making training more manageable while improving probability distribution estimation accuracy.

The patent uses parameter changes by adjusting the group number G in group convolution and the kernel sizes in different stages. By optimizing these parameters, the network achieves effective feature extraction with controlled complexity. The ability to independently tune group numbers and kernel sizes allows flexible adjustment of the trade-off between feature extraction capability and network complexity, facilitating easier training compared to fixed-structure deep networks.