Feature Data Compression by Channel Pruning for Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding standards face challenges in efficiently compressing multi-dimensional data, particularly for applications requiring machine consumption rather than human consumption, such as object recognition tasks, due to power and computational constraints at acquisition devices.
Innovation Solution
The method involves pruning channels from a tensor that do not satisfy a certain condition, signaling the pruned channels, and padding values to reconstruct the tensor, optimizing the compression of feature data for efficient distribution and analysis across a network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all channels in the tensor are transmitted, then complete feature data is available for accurate object recognition, but bandwidth and transmission time increase significantly
Solution Approach 1:
The patent extracts only the essential channels from the tensor based on importance metrics, removing redundant information. The encoder identifies and transmits only the most significant channels while discarding less important ones, achieving compression without significantly compromising recognition accuracy.
Solution Approach 2:
The patent changes the parameter of channel selection by using importance metrics to dynamically determine which channels to transmit. Instead of transmitting all channels uniformly, the system adapts the transmission set based on calculated importance values, optimizing the balance between accuracy and bandwidth usage.
2Quantity of substance
If feature data is compressed by removing channels, then bandwidth requirements are reduced, but information loss may occur
Solution Approach 1:
The patent introduces an importance metric as an intermediary that evaluates each channel's contribution to the overall feature representation. This intermediary assessment allows the system to make informed decisions about which channels to retain and which to discard, minimizing information loss while achieving compression.
Solution Approach 2:
The patent performs preliminary analysis of channel importance before transmission, calculating metrics that predict which channels are most valuable. This advance assessment allows the system to pre-select the optimal subset of channels to transmit, ensuring that the most important information is preserved while reducing overall data volume.
3Productivity
If channel pruning is applied to reduce data size, then transmission efficiency improves, but computational complexity increases due to channel selection and padding operations
Solution Approach 1:
The patent segments the channel selection process into distinct, manageable steps: calculating importance metrics, ranking channels, selecting the top subset, and applying padding. This segmentation allows each step to be optimized independently and makes the overall process more tractable and efficient.
Data Source
AI summary
A device may be configured to compress feature data according to one or more of the techniques described herein.