Two-Stream UNET Segmentation for Small Image Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deeper and larger Convolutional Neural Networks (CNNs) for image segmentation requires high-quality and large-scale datasets, which are expensive to acquire, especially in applications like medical image analysis and defect inspection, leading to lower accuracy when using smaller datasets.
Innovation Solution
A two-stream UNET architecture is employed for image segmentation, utilizing a spatial stream for pixel intensity values and a vector stream for gradient vector flow (GVF) to generate spatial and field feature maps, which are fused to improve segmentation accuracy on small datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deeper and larger CNNs are used for image segmentation, then segmentation accuracy is improved, but dataset size requirements increase and acquisition cost increases
Solution Approach 1:
The patent divides the CNN architecture into two separate streams: a spatial stream that processes spatial information and a feature stream that processes feature information. This segmentation allows each stream to be optimized independently, achieving high segmentation accuracy with smaller datasets by distributing the learning task across specialized pathways rather than requiring a single large network to learn all features simultaneously.
Solution Approach 2:
The patent introduces a dual-stream architecture that adds a dimensional aspect to feature processing by separating spatial and feature dimensions. The spatial stream handles spatial relationships while the feature stream handles semantic features, effectively transforming a single-dimension learning problem into a multi-dimensional processing framework that reduces data hunger.
2Measurement precision
If deeper and larger CNNs are used for image segmentation, then segmentation accuracy is improved, but acquisition cost increases
Solution Approach 1:
By segmenting the network into two streams with different功能的, the patent reduces the overall data acquisition burden. Each stream can be trained with smaller, more targeted datasets specific to its function, rather than requiring one massive dataset to train a single large network, thereby reducing acquisition costs while maintaining high accuracy.
Solution Approach 2:
The patent changes the architectural parameters of the CNN by implementing a two-stream structure with different numbers of layers and filters in each stream. This parameter diversification allows the system to achieve high accuracy without uniformly increasing the scale of the entire network, thus reducing the need for expensive large-scale data acquisition.
3Ease of manufacture
If smaller datasets are used for training, then data acquisition cost decreases, but segmentation accuracy deteriorates
Solution Approach 1:
The dual-stream architecture segments the learning task so that each stream specializes in specific aspects of image segmentation. This specialization allows each stream to achieve high performance with smaller, more focused datasets, overcoming the accuracy deterioration that typically occurs when using smaller datasets in conventional single-stream architectures.
Solution Approach 2:
The two-stream UNET architecture provides multi-functionality by handling both spatial processing and feature extraction through separate streams. This universal design allows the model to effectively learn from smaller datasets by distributing learning responsibilities across multiple functional pathways, maintaining high segmentation accuracy without requiring large amounts of training data.
Data Source
AI summary
Methods, apparatus, and systems are disclosed for semantic image segmentation using small datasets. An example apparatus includes at least one memory, machine readable instructions, and processor circuitry to at least one of instantiate or execute the machine readable instructions to identify a gradient vector flow associated with the input image, generate a spatial feature map based on pixels of the input image using a two-stream neural network architecture, generate a field feature map based on the gradient vector flow using the two-stream neural network architecture, fuse the spatial feature map and the field feature map, and output a segmented image of the input image based on the fused feature map.