Discriminative Path Tensor Segmentation for CNN Resource Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Semantic segmentation of image data using convolutional neural networks (CNNs) requires excessive calculation and memory resources, particularly when restoring image data to original resolution, which is unsustainable on embedded processing units without specialized hardware like GPUs.
Innovation Solution
The method introduces a discriminative path between the encoder and decoder paths in CNNs, dividing the input tensor into slice tensors and connecting them to maintain dimensionality, reducing resource usage and preventing gradient vanishing, allowing for efficient localization and classification on various hardware platforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a decoder component or upsampling component is added to restore image data to original resolution, then localization precision is improved, but calculation resource demand increases exponentially
Solution Approach 1:
The input tensor is divided into multiple slice tensors along the channel dimension. Each slice tensor is processed independently through the decoder path, allowing parallel computation and reducing the computational burden on any single processing unit. This segmentation enables the system to handle high-resolution image data without requiring exponential calculation resources.
2Manufacturing precision
If semantic segmentation is performed on a pixel basis, then manufacturing precision is improved, but memory bandwidth and memory space requirements increase
Solution Approach 1:
The input tensor is segmented into multiple slice tensors that can be processed independently. This segmentation reduces the amount of data that needs to be held in memory simultaneously, as each slice can be processed separately and results can be accumulated. This approach enables pixel-level segmentation while managing memory bandwidth and space requirements.
3Measurement precision
If the network processes high-resolution image data, then localization accuracy is improved, but memory bandwidth requirements increase
Solution Approach 1:
The high-resolution image data is processed by dividing the input tensor into multiple smaller slice tensors. Each slice is processed independently through the network, reducing the peak memory bandwidth requirements compared to processing the entire high-resolution tensor at once. The results from all slices are then combined to produce the final segmentation map.
Data Source
AI summary
Method for the calculation resource-saving semantic segmentation of image data of an imaging sensor with the aid of an artificial neural network, in particular, of a convolutional neural network, the artificial neural network including an encoder path, a decoder path, the encoder path transitioning into the decoder path, the transition taking place via a discriminative path, the following steps taking place in the discriminative path: dividing an input tensor as a function of a division function into at least one first slice tensor and at least one second slice tensor, the input tensor originating from the encoder path; connecting the at least one first slice tensor to the at least one second slice tensor as a function of a connection function in order to obtain a class tensor; and outputting the class tensor to the decoder path of the neural network.