Dual Neural Network Feature Map Fusion for Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The fully convolutional neural network lacks the specification of positional relations between elements in an image, which affects the accuracy of image recognition, and widening the recognition area to specify these relations deteriorates the accuracy.
Innovation Solution
A processing device utilizing two neural networks, one for wide-area recognition and one for narrow-area recognition, with a combiner to integrate their outputs, and an enlarger to adjust feature map sizes, all without fully connected layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a fully convolutional neural network is used for image processing, then the network can process images without fully connected layers, but the positional relation of elements in the image cannot be specified
Solution Approach 1:
The system segments the image processing task into two separate neural networks: a first neural network that processes the entire image to capture global positional relations, and a second neural network that processes a cropped local region to capture detailed features. This segmentation allows each network to specialize in different aspects of the problem, resolving the contradiction between maintaining simple convolutional structures and specifying positional relations.
Solution Approach 2:
The solution embeds one neural network's output within another's processing pipeline. The first neural network processes the full image and generates feature maps that are then cropped and fed into the second neural network. This nested structure allows the system to maintain the simplicity of convolutional networks while specifying positional relations through the hierarchical processing of nested feature maps.
2Measurement precision
If the recognition area is widened to specify positional relations, then positional information can be obtained, but the accuracy of detailed recognition deteriorates
Solution Approach 1:
The system divides the recognition task into two segments: global recognition of positional relations and local recognition of detailed features. By processing the entire image in the first network and a cropped local region in the second network, the system can simultaneously specify positional relations and maintain detailed recognition accuracy without compromise.
Solution Approach 2:
Different regions of the processing system are assigned different qualities and functions. The first neural network is optimized for global contextual understanding and positional relation specification, while the second neural network is optimized for local detailed feature recognition. This local quality differentiation allows each part to excel at its specific function, resolving the contradiction between global and local recognition accuracy.
Data Source
AI summary
A first processor executes processing of a first neural network on a target image to be processed and generates a first feature map having a size (1/m)×(1/n) times as large as the target image. An enlarger enlarges the first feature map generated in the first processor by n times. A second processor executes processing of a second neural network on the target image and generates a second feature map having a size (1/m) times as large as the target image. The combiner combines the first feature map enlarged by n times in the enlarger and the second feature map generated in the second processor.