CN-122023768-A - Data processing method and device for optimizing target detection model

CN122023768ACN 122023768 ACN122023768 ACN 122023768ACN-122023768-A

Abstract

The application discloses a data processing method and device for optimizing a target detection model. The method comprises the steps of obtaining to-be-processed scene data, performing incremental learning processing on the to-be-processed scene data based on first scene characteristics to obtain updated scene characteristic data, performing model updating processing on the updated scene characteristic data based on a first target detection model to obtain an updated target detection model, wherein the first target detection model is used for a target detection model in a previous target detection scene, and performing model performance evaluation processing on the updated target detection model to obtain a second target detection model. The learning and updating of the target detection model in the new scene are realized by gradually learning and updating the data in different scenes, and the instantaneity and the accuracy of target detection in the new scene and the environment are improved.

Inventors

ZHANG XIAOWEI
CHEN TONG
Luan Xinrui
LIU YING
YIN FEI
ZHANG YIMENG
DONG YUCAI
PENG YUTING
DONG WENTAO
Kong Zining
XIAO LONGBIN
DU JIAN
CUI WEI
LIN YUANYUAN
ZHANG SHITAI

Assignees

中国电子科技集团公司第十五研究所

Dates

Publication Date: 20260512
Application Date: 20260114

Claims (10)

1. A data processing method for optimization of a target detection model, comprising: acquiring to-be-processed scene data, wherein the to-be-processed scene data is data for representing the perceived interaction of the target detection scene environment; performing incremental learning processing on the scene data to be processed based on first scene characteristics to obtain updated scene characteristic data, wherein the first scene characteristics are scene characteristics of a previous target detection scene; Performing model updating processing based on a first target detection model on the updated scene characteristic data to obtain an updated target detection model, wherein the first target detection model is a target detection model used in a previous target detection scene; And performing model performance evaluation processing on the updated target detection model to obtain a second target detection model, wherein the second target detection model is an optimized target detection model.
2. The data processing method according to claim 1, wherein performing a model update process based on the first object detection model on the updated scene feature data, to obtain an updated object detection model, comprises: extracting scene characteristics of the scene data to be processed to obtain the scene characteristic data to be processed; Performing judgment processing based on scene similarity on the scene feature data to be processed and the first scene feature data, If the scene similarity of the scene feature data to be processed and the first scene feature data is greater than or equal to a preset similarity threshold value, a first model updating parameter is obtained, and the first target detection model is subjected to model updating processing according to the updated scene feature data in combination with the first model updating parameter, so that the updated target detection model is obtained; and if the scene similarity of the scene feature data to be processed and the first scene feature data is smaller than a preset similarity threshold value, obtaining a second model updating parameter, and carrying out model updating processing on the first target detection model according to the updated scene feature data and the second model updating parameter to obtain the updated target detection model.
3. The data processing method according to claim 2, wherein performing a model update process based on the first object detection model on the updated scene feature data, to obtain an updated object detection model, comprises: Performing model optimization processing based on model update parameters on the first target detection model to obtain a scene optimization first target detection model; forward reasoning processing is carried out on the updated scene characteristic data based on the scene optimization first target detection model, so that a model updating training set is obtained; And constructing a joint loss function according to the model updating training set and the scene optimization first target detection model, and training the target detection model through the joint loss function to obtain the updated target detection model.
4. A data processing method according to claim 3, wherein constructing a joint loss function from the model update training set and the scene optimization first target detection model, training the target detection model by the joint loss function, and obtaining the updated target detection model comprises: Performing associated feature extraction processing on the scene optimization first target detection model to obtain associated feature map data, wherein the associated feature map data are feature-to-feature relationship map data used for representing the scene optimization first target detection model; Performing prediction response processing on the model updating training set based on the scene optimization first target detection model to obtain prediction response data; constructing a joint loss function of a target detection model according to the associated feature map data and the predicted response data; And carrying out model iterative training processing on the target detection model based on the joint loss function to obtain the updated target detection model, wherein the updated target detection model is used for representing that the iterative training meets the convergence rule of the preset model.
5. The data processing method according to claim 1, wherein acquiring scene data to be processed includes: Acquiring multi-mode scene data, wherein the multi-mode scene data are data used for representing target detection scene environments acquired by various sensors; Respectively carrying out feature extraction processing on the multi-mode scene data to obtain a plurality of mode scene feature data; And carrying out data fusion processing based on feature fusion on the feature data of the plurality of modal scenes to obtain the scene data to be processed.
6. The data processing method according to claim 1, wherein after performing model performance evaluation processing on the updated target detection model to obtain a second target detection model, the method further comprises: performing target detection processing for a preset time period on a target detection scene according to the second target detection model to obtain target detection result data; Performing behavior feature extraction processing on the target detection result data to obtain target behavior feature data, wherein the target behavior feature data are feature data used for representing target behaviors in a scene within a preset time period; performing scene anomaly detection processing on the target behavior characteristic data to obtain anomaly detection result data, wherein the anomaly detection result data is used for representing scene anomalies in a target detection scene; and performing on-line learning processing on the scene anomaly detection model on the anomaly detection result to obtain an updated anomaly detection model.
7. The data processing method according to claim 1, wherein after performing model performance evaluation processing on the updated target detection model to obtain a second target detection model, the method further comprises: Acquiring task data to be updated, wherein the task data to be updated is data for representing a task to be updated in a target detection scene; Extracting first task features from task data to be updated to obtain first task features to be updated, wherein the first task feature data to be updated is data for representing static task features; Extracting second task features from the task data to be updated to obtain second task features to be updated, wherein the second task feature data to be updated is data for representing dynamic task features; Respectively carrying out incremental learning processing based on the prior task characteristics on the first task characteristic to be updated and the second task characteristic to be updated to obtain updated task characteristic data; performing model updating processing based on the first task model on the updated task feature data to obtain an updated task model, wherein the task model can be a scene anomaly detection model; and performing model performance evaluation processing on the updated task model to obtain a second task model, wherein the second task model is a task model after task updating.
8. A data processing apparatus for object detection model optimization, comprising: The data acquisition module is used for acquiring to-be-processed scene data, wherein the to-be-processed scene data is data for representing the perceived interaction of the target detection scene environment; The incremental learning module is used for performing incremental learning processing on the scene data to be processed based on first scene characteristics to obtain updated scene characteristic data, wherein the first scene characteristics are scene characteristics of a previous target detection scene; The model updating module is used for carrying out model updating processing based on a first target detection model on the updated scene characteristic data to obtain an updated target detection model, wherein the first target detection model is a target detection model used in a previous target detection scene; The model evaluation module is used for performing model performance evaluation processing on the updated target detection model to obtain a second target detection model, wherein the second target detection model is an optimized target detection model.
9. A computer-readable storage medium storing computer instructions for causing the computer to execute the data processing method for object detection model optimization according to any one of claims 1 to 7.
10. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores a computer program executable by the at least one processor to cause the at least one processor to perform the data processing method for object detection model optimization of any of claims 1-7.

Description

Data processing method and device for optimizing target detection model Technical Field The application relates to the field of computers, in particular to a data processing method and device for optimizing a target detection model. Background In recent years, an object detection model is widely used in various fields to realize object detection in various scenes. However, in the prior art, the target detection model mainly collects labeling sample data, performs model training according to the sample data, is difficult to effectively detect and identify data to be detected, which is different from sample data in data distribution, and the model needs to be retrained from the beginning for new data, so that the model training time is long and the efficiency is low. Military scenes have unique challenges, such as changeable environments, diversified targets, high real-time requirements and the like, a large amount of real-time data is usually required to be processed by a military target detection task, when an unmanned plane or a robot executes the task in the military scene, adaptive target detection, subsequent decision making and the like are required to be realized aiming at different scene data, and model detection in the prior art is difficult to realize. The application therefore proposes a data processing method and device for continuous optimization of a target detection model. Disclosure of Invention The application mainly aims to provide a data processing method and device for optimizing a target detection model, so as to solve one or more of the technical problems and achieve the technical effect of improving the real-time performance and accuracy of target detection. To achieve the above object, a first aspect of the present application proposes a data processing method for object detection model optimization, including: acquiring to-be-processed scene data, wherein the to-be-processed scene data is data for representing the perceived interaction of the target detection scene environment; performing incremental learning processing on the scene data to be processed based on first scene characteristics to obtain updated scene characteristic data, wherein the first scene characteristics are scene characteristics of a previous target detection scene; Performing model updating processing based on a first target detection model on the updated scene characteristic data to obtain an updated target detection model, wherein the first target detection model is a target detection model used in a previous target detection scene; And performing model performance evaluation processing on the updated target detection model to obtain a second target detection model, wherein the second target detection model is an optimized target detection model. Further, performing a model update process based on the first object detection model on the updated scene feature data, where obtaining an updated object detection model includes: extracting scene characteristics of the scene data to be processed to obtain the scene characteristic data to be processed; Performing judgment processing based on scene similarity on the scene feature data to be processed and the first scene feature data, If the scene similarity of the scene feature data to be processed and the first scene feature data is greater than or equal to a preset similarity threshold value, a first model updating parameter is obtained, and the first target detection model is subjected to model updating processing according to the updated scene feature data in combination with the first model updating parameter, so that the updated target detection model is obtained; and if the scene similarity of the scene feature data to be processed and the first scene feature data is smaller than a preset similarity threshold value, obtaining a second model updating parameter, and carrying out model updating processing on the first target detection model according to the updated scene feature data and the second model updating parameter to obtain the updated target detection model. Further, performing a model update process based on the first object detection model on the updated scene feature data, where obtaining an updated object detection model includes: Performing model optimization processing based on model update parameters on the first target detection model to obtain a scene optimization first target detection model; forward reasoning processing is carried out on the updated scene characteristic data based on the scene optimization first target detection model, so that a model updating training set is obtained; And constructing a joint loss function according to the model updating training set and the scene optimization first target detection model, and training the target detection model through the joint loss function to obtain the updated target detection model. Further, constructing a joint loss function according to the model update training set and the scene optimi