Jointly Trained Neural Compression for Data-Bus Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data compression methods for sensor data in multisensor systems are inefficient as they do not consider the specific requirements of subsequent machine learning processing, leading to suboptimal performance in object identification tasks.
Innovation Solution
A method involving a neural network with interconnected subnetworks is trained to optimize data compression based on the specific requirements of object identification, balancing compression ratio and identification accuracy by using a multi-objective optimization approach.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If standard compression methods (e.g., discrete cosine transformation, wavelet transformation, JPEG) are used to compress sensor data, then the data volume transmitted via the data bus is reduced, but the compression is unbiased compared to subsequent machine learning processing, causing relevant data to be lost and suboptimal object identification performance
Solution Approach 1:
The patent applies preliminary action by training the compression network in advance with knowledge of the object identification task. The compression network is pre-trained offline using annotated training data and a trained object identification network, so that it learns to compress data in a way that preserves features relevant for subsequent identification. This preliminary training ensures that when compression is applied during operation, it does not lose critical information needed for accurate object identification.
Solution Approach 2:
The patent implements feedback by using the object identification network to evaluate the quality of compressed data during the training phase. The compression network's parameters are adjusted based on feedback from the identification network's performance metrics (e.g., detection accuracy, false positive rates). This closed-loop feedback ensures that the compression algorithm learns to preserve features that are actually useful for the downstream identification task.
2Adaptability or versatility
If compression is applied to reduce data rate on the data bus, then more sensor devices can be connected to the data bus, but the data bus remains a significant bottleneck and the volume of data or data rate via the data bus may be reduced only by data compression
Solution Approach 1:
The system performs preliminary compression of sensor data before transmission over the data bus, reducing the data volume that needs to be transmitted. This allows more sensor devices to be connected to the data bus without overwhelming it, as each device transmits less data. The compression is performed offline-trained networks operating in real-time, enabling scalable sensor deployment.
3Productivity
If conventional compression methods are used, then the compression process is simple and fast, but they minimize general information criteria without considering the specific requirements of subsequent machine learning processing
Solution Approach 1:
The system performs preliminary training of the compression network offline using representative training data and the actual object identification network. This preliminary action allows the compression network to learn task-specific features that should be preserved, so that during actual operation, the compression is both efficient and informative for the downstream identification task.
Solution Approach 2:
The patent changes the parameters of the compression network during training to optimize for task-specific performance rather than general compression metrics. The network parameters (weights and biases) are adjusted through training to minimize a loss function that incorporates identification performance metrics, transforming the compression from a generic information-theoretic process to a task-aware process.
Data Source
AI summary
A method for configuring an object detection system includes providing annotated training data comprising image data with defined assignments to at least one object, and training a neural network with a first neural sub-network, which is provided to compress the image data. The first neural sub-network is connected to at least one further neural sub-network. The at least one further neural sub-network is configured to detect an object from the compressed training data. The first neural sub-network is parameterized in such a manner that the object is detected using the at least one further sub-network in a defined quality. The neural sub-networks are trained jointly.