Object Detection Model Retraining for False Positive Machine Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing models, such as convolutional neural networks, face challenges in accurately detecting and classifying objects prone to misclassification, leading to false positives and false negatives, which can result in inappropriate machine operational settings, potentially damaging items like clothes in washing machines.
Innovation Solution
A multi-pass training process is employed to generate improved image processing models by identifying and addressing 'difficult' data points, adding pseudo-classes to account for misclassification, and augmenting training data to enhance the models' precision and recall, allowing for accurate object detection and bi-class classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image processing models are used for object detection, then the machine can operate automatically, but the detection precision is poor leading to false positives and false negatives
Solution Approach 1:
The patent segments the classification task into multiple passes. The first pass identifies easy-to-classify samples, while the second pass focuses specifically on hard-to-classify samples that were misclassified in the first pass. This segmentation allows the model to progressively improve precision by addressing different subsets of data in separate training iterations.
Solution Approach 2:
The patent performs preliminary classification in the first pass to identify which samples are correctly classified and which are misclassified. This preliminary action enables the second pass to focus computational resources specifically on the problematic hard samples, improving overall detection precision without redundant processing of easy samples.
2Ease of manufacture
If the machine learning model classifies all samples uniformly, then the processing is simple, but hard samples cause consistent misclassification
Solution Approach 1:
The patent introduces dynamic adaptability by making the training process responsive to classification performance. The model automatically identifies hard samples through misclassification detection and adjusts its focus in subsequent training passes. This dynamic approach allows the system to adapt its processing strategy based on the difficulty of different samples rather than treating all samples uniformly.
Solution Approach 2:
The patent implements feedback mechanisms where the classification results of the first pass inform the training strategy of the second pass. Misclassified samples from the first pass are fed back into the training dataset for specialized attention in the second pass, creating a closed-loop system that continuously improves based on performance feedback.
3Productivity
If the model uses standard training data, then the training process is efficient, but it cannot distinguish between similar object types prone to misclassification
Solution Approach 1:
The patent performs preliminary identification of hard samples through the first classification pass before conducting specialized training. This preliminary action separates easy and hard samples, allowing the second pass to concentrate training efforts on distinguishing similar object types without wasting computational resources on already-correctly-classified samples.
Solution Approach 2:
The patent applies local quality improvement by providing differentiated training attention to different sample types. Hard samples that are prone to misclassification receive enhanced focus and specialized processing in the second pass, while easy samples maintain their standard processing. This localized quality enhancement improves overall precision without uniformly increasing complexity across all samples.
Data Source
AI summary
System and method for controlling a machine, including: receiving a first image processing model trained to classify an input image into a first class for images containing at least one object of a first type or a second class for images not containing an object of the first type; identifying a subset of inference results that are false positive results; generating a set of new training data from the first set of images, including augmenting an image in the first set of images to obtain a respective plurality of images and labeling the respective plurality of images as containing at least one object of a pseudo first class; training a second image processing model to classify an input image into the first class, the second class, and the first pseudo class; and modifying a device setting of a machine based on an inference result of the second image processing model.