CNN Block Feature Extraction via 1x1 Convolution and Elementwise Addition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CNNs face performance degradation in feature extraction due to increased computation and parameter amounts, especially with larger input images, leading to reduced feature map sizes and compromised image recognition and object detection performance.
Innovation Solution
The proposed method employs a learning and testing method using multiple blocks in a CNN, where a 1x1 convolution operation is applied followed by an MxM convolution operation, with elementwise addition and pooling, to generate feature maps while adjusting parameters through backpropagation, reducing computation and enhancing feature extraction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CNNs use larger filter kernels (MxM convolution) to extract features, then feature extraction accuracy is improved, but computation amount increases significantly
Solution Approach 1:
The patent segments the feature extraction process into multiple blocks, where each block contains a 1x1 convolution layer followed by an MxM convolution layer. This segmentation allows the computation to be distributed and managed in stages, reducing the computational burden of any single operation while maintaining overall extraction accuracy.
Solution Approach 2:
The 1x1 convolution layer performs a preliminary action by processing the input feature map before it enters the MxM convolution layer. This preliminary processing reduces the dimensionality and complexity of the data, thereby reducing the computation amount required for the subsequent MxM convolution while preserving important features for accurate extraction.
2Measurement precision
If conventional CNNs increase the number of blocks and layers to improve recognition performance, then image recognition accuracy is improved, but device complexity increases
Solution Approach 1:
The patent introduces dynamic elementwise addition that selectively integrates feature maps from different blocks based on their relevance and quality. This dynamic approach allows the network to adaptively adjust its complexity, using more blocks when needed and fewer when sufficient, thereby improving recognition accuracy without unnecessarily increasing device complexity.
Solution Approach 2:
Each block in the patent is designed as a universal module containing both 1x1 and MxM convolution layers that can serve multiple functions: feature extraction, dimensionality reduction, and feature map generation. This multi-functionality reduces overall network complexity by eliminating the need for separate specialized layers while maintaining high recognition accuracy.
3Measurement precision
If conventional CNNs process large input images to maintain detail quality, then feature extraction quality is improved, but processing time increases
Solution Approach 1:
The patent applies periodic pooling operations at strategic intervals between blocks to downsample feature maps. This periodic action maintains feature extraction quality by preserving important spatial information while reducing the image size at regular intervals, thereby decreasing processing time for subsequent operations without losing critical details.
Solution Approach 2:
The 1x1 convolution layers perform preliminary dimensionality reduction before the expensive MxM convolution operations. This preliminary action processes large input images more efficiently by reducing their size early in the pipeline, maintaining feature quality while significantly reducing the processing time required for subsequent feature extraction steps.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A learning method for extracting features from an input image by hardware optimization using n blocks in a convolutional neural network (CNN) is provided. The method includes steps of: a learning device instructing a first convolutional layer of a k-th block to elementwise add a (1_1)-st to a (k_1)-st feature maps or their processed feature maps, and instructing a second convolutional layer of the k-th block to generate a (k_2)-nd feature map; and feeding a pooled feature map, generated by pooling an ROI area on an (n_2)-nd feature map or its processed feature map, into a feature classifier; and instructing a loss layer to calculate losses by referring to outputs of the feature classifier and their corresponding GT. By optimizing hardware, CNN throughput can be improved, and the method becomes more appropriate for compact networks, mobile devices, and the like. Further, the method allows key performance index to be satisfied.