Convolutional Neural Network Segmentation for Image Morphing

The network is divided into three independent segments (down-sampling segment, densely connected segment, up-sampling segment) that process images in sequence. Each segment has specialized functions and operates independently, reducing computational overhead while maintaining feature extraction accuracy. The segmentation allows parallel processing within each segment without the need for complex inter-layer connections.

The patent extracts and removes redundant convolutional kernels and interconnected layers from the conventional CNN structure. By taking out only the essential processing components needed for style transfer and eliminating unnecessary complexity, the network achieves faster processing speed while preserving the core functionality of feature extraction and transformation.

Different segments of the network are assigned different levels of processing complexity appropriate to their specific functions. The down-sampling segment uses fewer kernels focused on structural features, the densely connected segment uses moderate kernels for style features, and the up-sampling segment uses fewer kernels for reconstruction. This local optimization of kernel density maintains comprehensive feature coverage while reducing total operation count.

The patent applies partial action by using sufficient but not excessive convolutional kernels in each segment. Rather than deploying excessive kernels throughout the entire network, the design uses just enough kernels in each specialized segment to achieve the required feature coverage, eliminating redundant operations and improving overall productivity.

The network architecture is segmented into three distinct functional blocks with clear boundaries. Each segment performs specific processing tasks (down-sampling, dense connection, up-sampling) and passes results to the next segment. This segmentation reduces network complexity by eliminating the need for dense interconnections between all layers while maintaining the capability for deep feature extraction through the sequential processing pipeline.

The patent transitions from the traditional horizontal dimension of many interconnected layers to a vertical dimension of three stacked segments. Each segment processes features at different scales and representations, effectively adding a dimensional aspect to feature extraction. This dimensional change allows deep feature extraction without requiring a large number of interconnected layers, thereby reducing network complexity.

By segmenting the network into distinct processing stages with clear input-output interfaces, the patent eliminates overlapping transactions between layers. Each segment processes its assigned data independently and passes results forward, ensuring that no computation is performed more than once. This segmentation maintains processing completeness while eliminating redundant operations, thereby improving operation efficiency.

The segmented architecture ensures continuous useful action by passing processed data sequentially from one segment to the next without backtracking or reprocessing. Each segment performs its designated function completely before passing control to the next segment, eliminating gaps or overlaps in the processing flow. This continuous action ensures comprehensive processing while maximizing operation efficiency by avoiding redundant computations.