CNN Mask Channel Convolution for Redundant Background Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Convolutional Neural Networks (CNNs) are compute-intensive due to redundant computations, particularly in background areas with complex details, and existing methods to reduce this intensity either require additional compute overhead or are limited in applicability.
Innovation Solution
The use of mask channels in input feature maps to mask pixels in CNNs, allowing for reduced redundant computations by generating masked feature channels and performing convolutions only on meaningful areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolution operations are performed for every pixel in input images and feature maps, then comprehensive feature extraction is achieved, but redundant computations occur in background areas
Solution Approach 1:
The patent segments the feature map processing into two distinct paths: a first convolutional layer processes the entire input feature map to generate comprehensive features, while a second convolutional layer processes only the selected regions identified by the mask channel. This segmentation allows the system to maintain comprehensive feature extraction capability while avoiding redundant computations in background areas, directly resolving the contradiction between feature extraction completeness and computation efficiency.
Solution Approach 2:
The patent applies local quality by making different parts of the feature map undergo different processing. The mask channel identifies regions requiring detailed processing (foreground objects) versus regions that can be processed with reduced computation (background areas). The second convolutional layer applies full processing only to selected regions while using results from the first layer for other areas, optimizing computational resources based on local importance.
2Productivity
If halting score computation is added to identify redundant pixels, then computation reduction is achieved, but additional matrix multiplication layers increase function call overhead
Solution Approach 1:
The patent merges the masking function into the existing convolutional layer structure by adding a mask channel to the standard convolutional layers. Instead of creating separate halting score computation layers, the mask channel is integrated into the input feature maps and processed together with the feature channels through the same convolution operations. This merging eliminates additional function call overhead while achieving computation reduction.
Solution Approach 2:
The mask channel serves multiple functions simultaneously: it acts as both an input feature and a control signal for computation reduction. The same mask channel is used to guide both the first and second convolutional layers, enabling a single additional channel to provide multi-functional benefits for computation optimization across different processing paths without requiring separate control mechanisms.
3Productivity
If background areas are omitted from computation, then computational efficiency is improved, but background areas with complicated details cannot be properly processed
Solution Approach 1:
The patent performs preliminary action by having the first convolutional layer process the entire input feature map before the second convolutional layer selectively processes specific regions. This preliminary full processing ensures that background areas with complicated details receive adequate initial feature extraction, and the results from this first pass are then utilized in regions where the mask channel indicates reduced computation is appropriate, maintaining accuracy while improving efficiency.
Solution Approach 2:
The first convolutional layer acts as an intermediary that bridges the gap between full processing and selective processing. It processes all regions including complex backgrounds to generate intermediate feature maps, which then serve as input for the second convolutional layer. This intermediary processing ensures that even regions ultimately processed with reduced computation have benefited from initial comprehensive feature extraction, maintaining reliability for complex backgrounds.
Data Source
AI summary
An object is to provide an information processing apparatus capable of reducing redundant computation in CNN. An information processing apparatus according to the present disclosure includes at least one memory configured to store an instruction, and at least one processor configured to execute the instruction to use mask channel in input feature maps to mask pixels of feature channels in the input feature maps and to generate masked feature channels, and perform a convolution operation between the masked feature channels and convolution kernel to generate output feature maps.