# US 2025/0095353 A1 · Method for detecting cigarette appearance defect based on variational bayesian inference

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0095353 A1
- Document ID: US20250095353A1
- Status: Published application
- Kind: A1
- Application number: 18/570,078
- Filed: 2023-02-24
- Published: 2025-03-20
- Earliest claimed: 2023-02-03
- Assignees: Hongyunhonghe Tobacco (Group) Co., Ltd
- Applicants: Hongyunhonghe Tobacco (Group) Co., Ltd
- Inventors: Xianzhou LV; Lin Qi; Yuxiang CUI; Dan LIN; Faqing LV; Liang CHENG; Lu YANG; Bing LIU; Meilin YI; Yunmei WANG; Yuedong QIAN; Shichae WU; Yunyu GONG
- Web page: https://patentgrep.com/patent/US20250095353A1

## CPC

- G06V 10/84: using probabilistic graphical models from image or video features, e.g. Markov models or Bayesian networks
- G06F 18/00: Pattern recognition
- G06N 3/042: Knowledge-based neural networks; Logical representations of neural networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 5/04: Inference or reasoning models
- G06T 7/0004: {Industrial image inspection}
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/766: using regression, e.g. by projecting features on hyperplanes
- G06V 10/7715: {Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods}
- G06V 10/82: using neural networks
- G06T 2207/20076: Probabilistic image processing
- G06T 2207/20081: Training; Learning
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T 2207/30108: Industrial image inspection
- Y02P 90/30: Computing systems specially adapted for manufacturing
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- Y02P: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS

## Abstract

The present application discloses a method for detecting cigarette appearance defects based on variational Bayesian inference. In a cigarette appearance defect detection scenario, in order to solve the problem that a current point estimation-based machine learning algorithm leads to an overconfidence decision in a data scarcity area, the present application proposes a variational inference-based Bayesian method to improve a SSD backbone network, preferably including calculating a real posterior probability by the variational inference method, replacing a point estimation mechanism of a convolution layer weight in the original backbone network by using a form of probability distribution, then detecting the improved extracted features by an SSD target detection algorithm, and furthermore, accurately determining whether the appearance of the cigarette is defective and the type and location of the defect. The present application can improve the accuracy of defect detection without losing the processing speed, not only saving a large amount of manpower and material resources, but also effectively providing reliable technical support for finding the root causes of defect types. Thus, it achieves intelligent quality risk management and control in a cigarette production workshop.

## Claims

### Claim 1

A method for detecting cigarette appearance defects based on variational Bayesian inference, characterized by comprising:
pre-constructing a cigarette appearance defect detection model based on the variational Bayesian inference, wherein the cigarette appearance defect detection model is an SSD target detection model, and the cigarette appearance defect detection model comprises an improved backbone network based on the variational Bayesian inference method;
collecting appearance images of cigarettes assembled and cut in a cigarette production line as data to be detected; and
performing real-time inference on the input cigarette appearance images by the cigarette appearance defect detection model to obtain a cigarette appearance defect detection result.

### Claim 2 (depends on 1)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 1, characterized in that the improved backbone network based on the variational Bayesian inference method comprises calculating a posterior probability by variational inference, and replacing a point estimation mechanism of a convolution layer weight in a backbone network of an original SSD target detection model by probability distribution.

### Claim 3 (depends on 1)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 1, characterized in that the training process of the cigarette appearance defect detection model comprises:
establishing a cigarette appearance defect detection data set as an input image data set;
extracting features by the improved backbone network by the variational Bayesian inference method to obtain a plurality of feature maps;
extracting information about the feature maps by a default box of each layer;
generating an offset of a detection box by a locator according to each piece of default box information, and comparing the offset with a ground truth to generate a locating loss;
generating a category label by a classifier according to each piece of default box information, and comparing the classified category with the ground truth to generate a classification loss;
integrating the classification loss and the locating loss to generate a total loss of the model and perform back propagation; and
performing repeated iteration by the image data set until the set number of training is reached and an optimal model parameter is saved.

### Claim 4 (depends on 3)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 3, characterized by further comprising, before performing the feature extraction;
inputting the cigarette appearance defect images and the ground truth, including a preset label and a bounding box; and
calculating an IOU related to the default box, and converting the ground truth into the form of default box.

### Claim 5 (depends on 1)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 1, characterized in that the performing real-time inference on the input cigarette appearance images by the cigarette appearance defect detection model comprises:
extracting feature maps based on the improved backbone network extracting information about the feature maps by a default box of each layer;
generating an offset of a detection box by a locator according to each piece of default box information; generating a category label by a classifier according to each piece of default box information;
eliminating invalid detection boxes and removing redundant detection boxes based on the detection box offset and the category label; and
outputting a final prediction box and a corresponding category label.

### Claim 6 (depends on 5)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 5, characterized in that the eliminating invalid detection boxes comprises eliminating detection boxes with labels as background.

### Claim 7 (depends on 5)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 5, characterized in that the removing redundant detection boxes comprises removing redundant detection boxes by non-maximum suppression.

### Claim 8 (depends on 1)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 1, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 9 (depends on 2)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 2, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 10 (depends on 3)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 3, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 11 (depends on 4)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 4, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 12 (depends on 5)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 5, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 13 (depends on 6)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 6, characterized in that the backbone network is a pre-feature extraction network VGG16.

### Claim 14 (depends on 7)

The method for detecting cigarette appearance defects based on variational Bayesian inference according to claim 7, characterized in that the backbone network is a pre-feature extraction network VGG16.
