Infrared imaging system-oriented MRTD intelligent test method
By combining non-uniformity correction and Faster R-CNN deep learning techniques, a lightweight network model with multi-scale feature fusion and attention mechanism is constructed, which solves the accuracy and efficiency problems of traditional MRTD detection and achieves high accuracy and real-time intelligent MRTD detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH WEST INST OF TECHN PHYSICS
- Filing Date
- 2025-11-26
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional MRTD detection methods rely on human observation, which suffers from low detection accuracy, poor efficiency, weak repeatability, and high labor costs, making it difficult to meet the needs of efficient R&D and mass production of infrared imaging systems.
A hierarchical detection algorithm consisting of preprocessing, feature extraction, and intelligent decision-making is adopted. Combined with non-uniformity correction and Faster R-CNN deep learning techniques, a lightweight network model with multi-scale feature fusion and attention mechanism is constructed for intelligent MRTD detection.
It significantly improves detection accuracy and speed, reduces subjective errors, lowers computational resource requirements, and is suitable for engineering and industrial applications.
Smart Images
Figure CN121961986A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of infrared imaging system testing technology, and relates to an intelligent MRTD testing method for infrared imaging systems. Background Technology
[0002] Against the backdrop of rapid advancements in modern technology, infrared imaging systems, with their unique passive detection advantages, have become an indispensable core technology in fields such as national defense, industrial inspection, and intelligent security. As a passive detection method, infrared imaging systems capture the infrared energy radiated by the target itself and convert it into a visible image, possessing powerful detection capabilities and becoming a key piece of equipment for intelligence gathering and precision strikes in modern warfare. The minimum resolvable temperature difference (MRTD), as a core indicator for evaluating the performance of infrared imaging systems, directly determines the accuracy of target detection, identification, and positioning in complex environments. Traditional MRTD detection technology mainly relies on human observation and subjective judgment. Typically, several trained professionals directly observe the target image of the system under test. At a certain spatial frequency, when the four-bar target pattern (7:1 aspect ratio) can be distinguished, the temperature difference between the target and the background is the MRTD at that spatial frequency. This detection method has inherent defects such as low detection accuracy, poor efficiency, and weak repeatability. It is not only affected by subjective factors such as the individual experience of the inspectors and visual fatigue, leading to large deviations in the detection results, but also has high labor costs and long detection cycles when facing massive detection tasks, which seriously restricts the development progress and mass production efficiency of infrared imaging systems.
[0003] In recent years, the use of deep learning technology to achieve intelligent detection of MRTD has become a research hotspot. In the paper "Research on Objective Testing Method of MRTD Based on Machine Learning" (Ji Ran, Xiao Maosen, Li Shuo, et al. Research on Objective Testing Method of MRTD Based on Machine Learning [J]. Systems Engineering and Electronics, 2024, 46(10): 3265-3270.), the authors used two intelligent testing methods, Support Vector Machine (SVM) and Convolutional Neural Network (CNN), for detection. This method uses infrared images as the input of the neural network and distinguishable (1) and indistinguishable (0) as the output of the neural network. The results show that both models have achieved good results in terms of accuracy and real-time performance, and can solve the subjective problem of traditional MRTD, verifying the feasibility of intelligent detection method based on deep learning technology. However, constructing an intelligent detection method with high accuracy and easy engineering implementation is still an urgent problem to be solved. Summary of the Invention
[0004] (a) Purpose of the invention The purpose of this invention is to propose an intelligent MRTD testing method for infrared imaging systems. It adopts a hierarchical detection algorithm model structure of "preprocessing-feature extraction-intelligent decision-making", combines image preprocessing techniques such as non-uniformity correction with deep learning technology based on Faster R-CNN, takes into account the lightweight design requirements of the network model, achieves high accuracy and intelligent detection of MRTD, and ensures real-time detection, while reducing repetitive work and subjective interference in traditional testing methods.
[0005] (II) Technical Solution To address the aforementioned technical problems, this invention provides an intelligent MRTD testing method for infrared imaging systems, comprising the following steps: Step 1: Build the MRTD test hardware platform, adjust parameters such as blackbody temperature and target spatial frequency, and under the condition that the spatial frequency remains unchanged, collect the original four-bar target image data of infrared thermal imagers of the same model at different temperatures at the same spatial frequency. Step 2: The acquired raw image data is cropped to obtain only the important regions of the four-bar target image. That is, the 1280*1024 image is cropped into a 300*300 image with the four-bar target image as the center. The cropped image data is then preprocessed using algorithms such as non-uniformity correction to improve image quality. Step 3: In order to make full use of the limited data samples, data augmentation methods are used to expand the image, that is, to perform operations such as rotation, zoom, cropping, and offset on the image to expand the data samples. Step 4: The images are labeled, and the labels are divided into clearly distinguishable (label value is the letter Y), MRTD threshold (label value is the letter M), and indistinguishable (label value is the letter N). Then the labeled images are divided into training set, test set and validation set. Step 5: Construct the MRTD intelligent detection model. Based on the Faster R-CNN algorithm, make targeted improvements by introducing a multi-scale feature fusion module and an attention mechanism module to enhance the detection accuracy of the algorithm. At the same time, use depthwise separable convolution to make lightweight improvements to the network model, reduce memory consumption, and improve the real-time performance of detection. Step 6: Train the MRTD intelligent detection model using image data and label values from the training set. Test the accuracy of the network model using image data and label values from the test set until the accuracy reaches 97% or higher. Finally, use the trained network model to predict the label values of the image data from the validation set and compare them with the label values of the validation set to determine the accuracy of the network model. The temperature difference value corresponding to the predicted label value N is the MRTD value of the infrared thermal imager at that spatial frequency.
[0006] (III) Beneficial Effects The MRTD intelligent testing method for infrared imaging systems provided by the above technical solution has the following beneficial effects: (1) This invention addresses the problem of image quality degradation caused by non-uniformity of infrared detectors by using non-uniformity correction technology to preprocess infrared images, which significantly improves the quality of infrared images and helps feature learning of subsequent network models, thereby improving detection accuracy and reliability. (2) This invention innovatively improves upon the Faster R-CNN model and constructs an end-to-end deep learning network model based on multi-scale feature fusion and attention mechanism, which effectively solves problems such as easy missed detection of small targets and interference from complex backgrounds. Compared with traditional methods that require multiple steps such as image enhancement, edge extraction, and feature matching, this model directly maps the original infrared image data to the MRTD detection results, reducing the error transmission in intermediate links and effectively improving the detection accuracy. (3) In order to solve the problem of limited computing resources faced by deep learning models when deployed on edge devices, this invention proposes to use depthwise separable convolution technology to design a lightweight network model, further reducing the memory occupation of the network model. Through two steps of depthwise convolution and pointwise convolution, the computational load of the model is greatly reduced, the running efficiency of the model is improved, and the engineering and industrial application of the MRTD testing method is facilitated. Attached Figure Description
[0007] Figure 1 This is a flowchart of an MRTD intelligent testing method for infrared imaging systems.
[0008] Figure 2 This is the raw output image of the infrared imaging system.
[0009] Figure 3 This is the image after preprocessing.
[0010] Figure 4 This is a structural diagram of the MRTD detection algorithm.
[0011] Figure 5 This refers to the network model training process.
[0012] Figure 6 This is a graph showing the MRTD detection results. Detailed Implementation
[0013] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0014] Combination Figure 1This invention presents an intelligent MRTD testing method for infrared imaging systems. It combines image preprocessing techniques with deep learning technology based on Faster R-CNN to achieve intelligent MRTD testing, while employing a lightweight design to reduce the memory footprint of the network model. Based on deep learning algorithms such as YOLO and Faster R-CNN, through feature extraction, anchor box mechanisms, and region proposal networks, it can achieve rapid localization and accurate classification of targets in images. Its detection accuracy in various scenarios far surpasses traditional methods. MRTD testing is essentially image data processing, suitable for deep learning technology frameworks. It leverages the powerful feature extraction and pattern recognition capabilities of deep learning to intelligently process infrared MRTD test images, effectively avoiding errors caused by subjective human observation. The specific testing steps are as follows: Step 1: Build the MRTD test hardware platform and collect image data according to the data acquisition process. For the same model of infrared imaging system, set different temperature difference gradients and collect image data at ±0.4K, ±0.6K, ±0.8K, ±1.0K, ±2.0K, and ±3.0K as training and validation datasets for the network model. Collect image data at +0.5K, +0.7K, -0.45K, and -1.5K as test datasets for the network model.
[0015] In step 1 above, the acquired infrared imaging system image can be expressed as follows: Where N(x,y) represents the total information of the infrared image, X(x,y) represents the thermal radiation information of the target object captured by the infrared imaging system, Y(x,y) represents the thermal radiation information generated by impurities in the air and other interfering objects besides the target, and Z(x,y) represents the background information.
[0016] Step 2, as in 2 and Figure 3 As shown, in order to increase the proportion of effective information such as the four-bar target, the acquired original image data is cropped to obtain only the important regions of the four-bar target image. That is, the 1280*1024 image is cropped into a 300*300 image with the four-bar target image as the center, which contains complete information of the four-bar target and reduces the interference of background information. At the same time, algorithms such as non-uniformity correction are used to preprocess the cropped image data to improve image quality.
[0017] Step 3: In order to make full use of the limited data samples, data augmentation methods are used to expand the image, that is, to perform operations such as rotation, zoom, cropping, and offset on the image to expand the data samples.
[0018] Step 4: The images are labeled. The labels are divided into clearly distinguishable (label value is the letter Y), MRTD threshold (label value is the letter M), and indistinguishable (label value is the letter N). Then the labeled images are divided into training set, test set and validation set.
[0019] Step 5: Use the image data from the training and test sets as input to train the network model. The structure of the network model is as follows: Figure 6 As shown, this enables the network model to accurately classify images, predicting label values Y (corresponding to clearly distinguishable), M (corresponding to the MRTD threshold), and N (corresponding to indistinguishable). The predicted label values are compared with the true label values. If the accuracy of the comparison is lower than 97%, the training is repeated until the accuracy is greater than 97%.
[0020] In step 5 above, the performance of the network model is primarily evaluated using the average detection accuracy: Where mAP represents the average detection accuracy, and k represents the total number of target categories in the detection task. This represents the average accuracy of the i-th type of target.
[0021] In step 5 above, during the network model construction and training process, depthwise separable convolution is used to perform lightweight design of the network model, reducing the computational cost of the network model.
[0022] Step 6: Use the network model trained in Step 5 to intelligently interpret image data with unknown state. The main method is to use the image data in the validation set as the input of the trained network model, and predict the label value of the corresponding image through the analysis and calculation of the network model. Finally, the predicted label value is compared and verified with the real image label value. The temperature difference value corresponding to the image with the label value of N is the MRTD value of the infrared imaging system at that spatial frequency.
[0023] Based on the above technical solution, this invention combines image preprocessing technology with deep learning technology based on Faster R-CNN network to achieve intelligent testing of MRTD, effectively reducing the subjective error of testers, greatly improving the detection speed, and possessing high detection accuracy. Simultaneously, the network model is designed with lightweight features, which is beneficial for engineering applications. Specifically, it has the following significant characteristics: (1) Compared with traditional manual methods, the intelligent testing method of the present invention can effectively avoid problems such as subjective bias, low detection efficiency and insufficient accuracy, and significantly improve the detection accuracy and detection speed of MRTD.
[0024] (2) Compared with conventional neural networks, the test method of the present invention integrates image preprocessing technology and uses non-uniformity correction and other methods to process infrared images, which effectively improves image quality and helps feature learning of subsequent network models. At the same time, the accuracy of the network model based on Faster R-CNN in MRTD detection can reach more than 97%, while the accuracy of conventional neural network models is generally around 95%.
[0025] (3) The present invention uses depthwise separable convolution technology to design a lightweight network model, which effectively reduces the memory usage of the network model and further improves the speed and real-time performance of MRTD detection, which is conducive to promoting the engineering and industrial application of MRTD detection.
[0026] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A smart MRTD testing method for infrared imaging systems, characterized in that, Includes the following steps: Step 1: Build the MRTD test hardware platform, adjust the blackbody temperature and target spatial frequency parameters, and under the condition that the spatial frequency remains unchanged, collect the original four-bar target image data of infrared thermal imagers of the same model at different temperatures at the same spatial frequency. Step two: Crop the acquired raw image data to obtain only the important regions of the four target images; Step 3: Use data augmentation methods to augment the image; Step four: The images are labeled, and the labeled images are divided into training set, test set and validation set; Step 5: Construct the MRTD intelligent detection model. Based on the Faster R-CNN algorithm, introduce a multi-scale feature fusion module and an attention mechanism module, and use depthwise separable convolution to improve the model's lightweight nature. Step 6: Train the MRTD intelligent detection model using image data and label values from the training set. Test the model's accuracy using image data and label values from the test set. Finally, use the trained model to predict label values from image data in the validation set and compare them with the label values in the validation set to determine the model's accuracy. The temperature difference value corresponding to the predicted label value N is the MRTD value of the infrared thermal imager at that spatial frequency.
2. The MRTD intelligent testing method for infrared imaging systems as described in claim 1, characterized in that, In step two, the 1280*1024 image is cropped into a 300*300 image, centered on the four-bar target image. Non-uniformity correction is then used to preprocess the cropped image data to improve image quality.
3. The MRTD intelligent testing method for infrared imaging systems as described in claim 2, characterized in that, In step three, image augmentation processing involves rotating, zooming, cropping, and offsetting the image to augment the data sample.
4. The MRTD intelligent testing method for infrared imaging systems as described in claim 3, characterized in that, In step four, the labels are categorized as follows: Clearly distinguishable: label value is the letter Y; MRTD threshold: label value is the letter M; Indistinguishable: label value is the letter N.
5. The MRTD intelligent testing method for infrared imaging systems as described in claim 4, characterized in that, In step one, an MRTD test hardware platform is built, and image data is acquired according to the data acquisition process. For the same model of infrared thermal imager, different temperature difference gradients are set, and image data at ±0.4K, ±0.6K, ±0.8K, ±1.0K, ±2.0K, and ±3.0K are acquired as training and validation datasets for the network model, and image data at +0.5K, +0.7K, -0.45K, and -1.5K are acquired as test datasets for the network model.
6. The MRTD intelligent testing method for infrared imaging systems as described in claim 5, characterized in that, The images acquired in step one by the infrared thermal imager are shown below: Where N(x,y) represents the total information of the infrared image, X(x,y) represents the thermal radiation information of the target object captured by the infrared thermal imager, Y(x,y) represents the thermal radiation information generated by impurities in the air and other interfering objects besides the target, and Z(x,y) represents the background information.
7. The MRTD intelligent testing method for infrared imaging systems as described in claim 6, characterized in that, In step five, the model is trained using image data from the training and test sets as input, enabling the model to accurately classify images and predict label values Y, M, and N. The predicted label values are compared with the true label values. If the accuracy of the comparison is lower than 97%, the training is repeated until the accuracy is greater than 97%.
8. The MRTD intelligent testing method for infrared imaging systems as described in claim 7, characterized in that, In step five, the accuracy of the comparison is evaluated using the average detection accuracy: Where mAP represents the average detection accuracy, and k represents the total number of target categories in the detection task. This represents the average accuracy of the i-th type of target.
9. The MRTD intelligent testing method for infrared imaging systems as described in claim 8, characterized in that, In step six, the model trained in step five is used to intelligently interpret image data with unknown states. The image data in the validation set is used as the input to the trained network model. Through the analysis and calculation of the network model, the label value of the corresponding image is predicted. Finally, the predicted label value is compared and verified with the real image label value. The temperature difference value corresponding to the image with the label value of N is the MRTD value of the infrared imaging system at that spatial frequency.
10. An application of the MRTD intelligent testing method for infrared imaging systems based on any one of claims 1-9 in the testing of infrared imaging systems.