Distributed Image Classification via Convolutional Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning-based image classification methods rely heavily on cloud devices for computational tasks, leading to increased load and inefficiencies in data transmission and processing.
Innovation Solution
Distributing the image classification process between terminal devices and cloud devices, where terminal devices perform initial convolution operations and transmit intermediate results to cloud devices for further processing, thereby reducing the computational burden on the cloud and optimizing data transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the cloud device carries all computational duties in deep learning image classification, then the AI model can be fully trained and inferred, but the computational load and processing time on the cloud device increase significantly
Solution Approach 1:
The patent divides the image classification model into multiple segments: convolutional layers are deployed on terminal devices while fully connected layers remain on cloud devices. This segmentation allows local execution of feature extraction operations, reducing the time-consuming data transmission and cloud processing burden while maintaining overall classification accuracy.
2Reliability
If the cloud device carries all computational duties in deep learning image classification, then complete AI model processing is achieved, but the computational load on the cloud device increases
Solution Approach 1:
The patent segments the computational workload by deploying convolutional layers on terminal devices and retaining only fully connected layers on cloud devices. This distribution significantly reduces the computational load on cloud devices while preserving the accuracy benefits of complete model processing.
Solution Approach 2:
Terminal devices perform self-service by executing convolutional operations locally using their own processors. This eliminates the need to transmit raw images to the cloud for initial processing, reducing cloud computational requirements while maintaining classification quality.
3Reliability
If the terminal transmits the entire image to the cloud device for classification, then the cloud device can perform complete inference, but the network traffic and data transmission volume increase
Solution Approach 1:
The patent segments the processing workflow so that terminal devices perform convolutional feature extraction locally and only transmit intermediate feature maps to the cloud. This segmentation dramatically reduces the data transmission volume compared to sending complete high-resolution images, while still enabling accurate cloud-based classification.
4Productivity
If convolutional layers are performed on terminal devices, then processing efficiency improves and cloud load reduces, but the terminal device computational requirements increase
Solution Approach 1:
The patent extracts and removes the computationally intensive convolutional layers from the cloud-based model architecture, deploying them on terminal devices. This extraction improves overall efficiency by enabling local preprocessing while the cloud device handles only the lighter fully connected layers, effectively managing the complexity distribution.
Data Source
AI summary
A cloud device, a terminal device, and an image classification system and method includes the terminal reading an image to be classified; transmitting an instruction of image classification to the cloud device; determining whether the instruction of image classification comprises an intermediate operation result of the first partial operation; performing the second partial operation to obtain an image classification result, when the instruction of image classification comprises intermediate operation result of the first partial operation; and transmitting the image classification result to the terminal device; performing the first partial operation to obtain the intermediate operation result, when the intermediate operation result is not contained in the instruction of image classification; transmitting the intermediate operation result to the terminal device; and performing the second partial operation to obtain an image classification result.