CN-122023747-A - Image hidden danger identification parameter optimization method and system based on cascading model

CN122023747ACN 122023747 ACN122023747 ACN 122023747ACN-122023747-A

Abstract

The invention discloses a cascade model-based image hidden danger identification parameter optimization method and system, and relates to the technical field of data processing. The method comprises the steps of collecting a target image, carrying out hidden danger area identification to obtain a hidden danger candidate area set, carrying out fine-granularity identification on each hidden danger candidate area in the hidden danger candidate area set to obtain a plurality of identification hidden danger descriptions, introducing a standard hidden danger reference gallery, carrying out feature similarity calculation on the hidden danger candidate area set to obtain a plurality of comparison hidden danger descriptions, and generating a comprehensive hidden danger identification result according to the plurality of identification hidden danger descriptions and the plurality of comparison hidden danger descriptions, wherein before the hidden danger area is identified, parameter optimization is carried out on a large model and feature similarity calculation parameters of fine-granularity identification according to a scene of the target image. The invention effectively improves the identification accuracy and scene suitability of the hidden image trouble.

Inventors

JIA BO
Zhao Gouqiu
Ying Shengnan
LI XUESHENG
XIE BIFENG
FENG GUI
Lei zhuo
FANG ZHENG
SHEN CHEN
XU XIAOLING
ZHOU TINGTING

Assignees

浙江省应急管理科学研究院(浙江省安全生产技术检测检验中心、浙江省危险化学品登记中心)
杭州峰景科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The image hidden danger identification parameter optimization method based on the cascade model is characterized by comprising the following steps of: acquiring a target image, and identifying hidden danger areas to obtain a hidden danger candidate area set; carrying out fine-grained recognition on each hidden danger candidate region in the hidden danger candidate region set to obtain a plurality of recognition hidden danger descriptions; introducing a standard hidden danger reference gallery, and calculating feature similarity of the hidden danger candidate region set to obtain a plurality of comparison hidden danger descriptions; Generating a comprehensive hidden danger identification result according to the multiple identification hidden danger descriptions and the multiple comparison hidden danger descriptions; Before the hidden danger area is identified, parameter optimization is carried out on the large model identified in fine granularity and the feature similarity calculation parameters according to the scene of the target image.
2. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 1, wherein the steps of collecting a target image, identifying hidden danger areas, and obtaining a hidden danger candidate area set include: performing feature extraction on the target image by using a lightweight convolutional neural network to generate a preliminary feature map; And generating a plurality of hidden danger candidate region sets through a bounding box regression method based on the preliminary feature map.
3. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 1, wherein the step of performing fine-grained identification on each hidden danger candidate region in the hidden danger candidate region set to obtain a plurality of identification hidden danger descriptions includes: acquiring a sample hidden danger candidate region set and a sample identification hidden danger description set; training a hidden danger identification model by adopting the sample hidden danger candidate region set and the sample hidden danger identification description set based on a neural network; Inputting the images of the hidden trouble candidate areas into a hidden trouble recognition model, and outputting to obtain a plurality of recognition hidden trouble descriptions.
4. The method for optimizing image hidden danger identification parameters based on a cascade model of claim 3, wherein training the hidden danger identification model comprises: Collecting a sample hidden danger candidate region image set, and labeling to obtain a sample identification hidden danger description set; and taking the sample hidden danger candidate region image set as input, taking the sample hidden danger identification description set as supervision, training a hidden danger identification model until convergence, and generating a hidden danger identification model.
5. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 1, wherein the step of introducing a standard hidden danger reference gallery, and performing feature similarity calculation on the hidden danger candidate region set to obtain a plurality of comparison hidden danger descriptions comprises the steps of: constructing a standard hidden danger reference graph library covering typical hidden danger descriptions, and carrying out feature extraction and vectorization storage on each standard hidden danger reference graph; extracting features of each hidden danger candidate region to generate a plurality of candidate region feature vectors; Calculating the similarity of each candidate region feature vector and each standard hidden danger reference picture feature vector in the standard hidden danger reference picture library to obtain a plurality of comparison similarity sets; And respectively selecting hidden danger descriptions corresponding to the standard hidden danger reference diagrams with highest comparison similarity as the comparison hidden danger descriptions of the hidden danger candidate areas, and outputting a plurality of comparison hidden danger descriptions.
6. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 4, wherein the feature extraction is performed on each hidden danger candidate region to generate a plurality of candidate region feature vectors, comprising: processing each hidden danger candidate area image by using a feature extraction network; and extracting candidate region feature vectors from the output of the global pooling layer following the full connection layer or the last convolution layer of the feature extraction network to obtain a plurality of candidate region feature vectors.
7. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 1, wherein before identifying hidden danger areas, the method for optimizing parameters of a large model and feature similarity calculation parameters for fine-grained identification according to the scene of the target image comprises the following steps: Constructing a parameter optimization database, and storing mapping relations between different application scenes and optimal recognition parameter configuration, wherein the recognition parameters at least comprise confidence thresholds for hidden danger area recognition, attention weights for fine-granularity recognition areas and feature matching similarity thresholds; and inquiring the parameter optimization database according to the scene characteristics of the target image, and loading the corresponding optimal identification parameter configuration.
8. The image hidden danger identification parameter optimization method based on the cascading model according to claim 6, the method is characterized by further comprising the step of dynamically optimizing and adjusting the identification parameters, and comprising the following steps: testing to obtain an identification effect index of the current identification parameter, and calculating the deviation amount of the identification effect index and a target index, wherein the identification effect index comprises accuracy; Calculating the deviation rate according to the deviation amount, and calculating a parameter adjustment amount; And adjusting the identification parameters according to the parameter adjustment amount until convergence, and obtaining the optimal identification parameters of different application scenes.
9. The method for optimizing image hidden danger identification parameters based on a cascading model according to claim 1, wherein generating a comprehensive hidden danger identification result according to the plurality of identification hidden danger descriptions and the plurality of comparison hidden danger descriptions comprises: When the identification hidden trouble description of the candidate area is consistent with the comparison hidden trouble description, outputting and obtaining a comprehensive hidden trouble identification result; When the hidden danger identification description is inconsistent with the hidden danger comparison description, outputting two hidden danger descriptions as a comprehensive hidden danger identification result.
10. A cascade model-based image hidden danger identification parameter optimization system, characterized in that it is used for implementing the cascade model-based image hidden danger identification parameter optimization method according to any one of claims 1-8, and the system comprises: The hidden danger area preliminary screening module is used for collecting target images, identifying hidden danger areas and obtaining a hidden danger candidate area set; The hidden danger fine recognition module is used for carrying out fine granularity recognition on each hidden danger candidate region in the hidden danger candidate region set to obtain a plurality of recognition hidden danger descriptions; the feature comparison analysis module is used for introducing a standard hidden danger reference gallery, and carrying out feature similarity calculation on the hidden danger candidate region set to obtain a plurality of comparison hidden danger descriptions; the hidden danger comprehensive research and judgment module is used for generating a comprehensive hidden danger identification result according to the multiple identification hidden danger descriptions and the multiple comparison hidden danger descriptions; Before the hidden danger area is identified, parameter optimization is carried out on the large model identified in fine granularity and the feature similarity calculation parameters according to the scene of the target image.

Description

Image hidden danger identification parameter optimization method and system based on cascading model Technical Field The invention relates to the technical field of data processing, in particular to a cascade model-based image hidden danger identification parameter optimization method and system. Background With the development of artificial intelligence technology, the picture AI hidden danger identification has remarkable application value in the fields of urban security, industrial manufacturing and the like, can identify hidden danger such as unsafe states of objects, unsafe behaviors of people and the like, and improves public safety level and production efficiency. However, in the prior art, a single model or a fixed structure is generally adopted, so that the method is difficult to be well adapted when the method is used for illumination change, shielding interference and cross-scene application of a complex scene, and further the identification accuracy and the scene suitability are poor. Disclosure of Invention The application provides a cascade model-based image hidden danger identification parameter optimization method and system, and aims to solve the technical problems of poor identification accuracy and poor scene suitability in the prior art. In view of the above problems, the application provides a cascade model-based image hidden danger identification parameter optimization method and system. In a first aspect, the present application provides a cascade model-based image hidden danger identification parameter optimization method, including: acquiring a target image, and identifying hidden danger areas to obtain a hidden danger candidate area set; carrying out fine-grained recognition on each hidden danger candidate region in the hidden danger candidate region set to obtain a plurality of recognition hidden danger descriptions; introducing a standard hidden danger reference gallery, and calculating feature similarity of the hidden danger candidate region set to obtain a plurality of comparison hidden danger descriptions; Generating a comprehensive hidden danger identification result according to the multiple identification hidden danger descriptions and the multiple comparison hidden danger descriptions; Before the hidden danger area is identified, parameter optimization is carried out on the large model identified in fine granularity and the feature similarity calculation parameters according to the scene of the target image. In a second aspect, the present application provides an image hidden danger identification parameter optimization system based on a cascading model, including: The hidden danger area preliminary screening module is used for collecting target images, identifying hidden danger areas and obtaining a hidden danger candidate area set; The hidden danger fine recognition module is used for carrying out fine granularity recognition on each hidden danger candidate region in the hidden danger candidate region set to obtain a plurality of recognition hidden danger descriptions; the feature comparison analysis module is used for introducing a standard hidden danger reference gallery, and carrying out feature similarity calculation on the hidden danger candidate region set to obtain a plurality of comparison hidden danger descriptions; the hidden danger comprehensive research and judgment module is used for generating a comprehensive hidden danger identification result according to the multiple identification hidden danger descriptions and the multiple comparison hidden danger descriptions; Before the hidden danger area is identified, parameter optimization is carried out on the large model identified in fine granularity and the feature similarity calculation parameters according to the scene of the target image. One or more technical schemes provided by the application have at least the following technical effects or advantages: The application provides an image hidden danger identification parameter optimization method and system based on a cascading model, which are used for accurately locking a suspected hidden danger range by collecting a target image and identifying a hidden danger candidate region, carrying out fine-granularity identification on the candidate region to finely distinguish hidden danger descriptions, then introducing a standard hidden danger reference graph library to carry out feature similarity calculation, enhancing the accuracy of feature matching, finally combining two types of identification results to generate a comprehensive hidden danger identification result, and optimizing relevant parameters according to a scene before identification, thereby effectively improving the accuracy of hidden danger identification and the scene suitability. Drawings In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is