Taper sleeve infrared thermal imaging target identification method based on improved yov5s and related equipment thereof

Abstract

The invention provides a taper sleeve infrared thermal imaging target identification method based on improved yov5s and related equipment thereof. The taper sleeve infrared thermal imaging target identification method comprises the following steps: acquiring to-be-detected infrared thermal imaging image data; inputting the to-be-detected infrared thermal imaging image data into a preset infrared thermal imaging detection model for reasoning processing to obtain a taper sleeve target candidate area; post-processing is carried out on the taper sleeve target candidate area, a taper sleeve target detection result is determined, and post-processing comprises confidence coefficient screening and non-maximum suppression. Through the steps of the method, rapid identification and stable output of the taper sleeve target can be realized in an infrared thermal imaging scene, the missed detection and false detection risks are reduced, and the method is suitable for an infrared target detection application scene with relatively high requirements on real-time performance and accuracy.

Description

Technical Field

[0001] This invention relates to the field of infrared thermal imaging recognition, and in particular to a method, apparatus, electronic device and storage medium for cone-shaped infrared thermal imaging target recognition based on an improved YOLOv5s. Background Technology

[0002] Infrared thermal imaging technology can form images based on the temperature difference between a target and its background. It can still acquire stable images in low-light, nighttime, and smoke-filled environments where visible light is limited, thus finding wide application in security monitoring, industrial inspection, and military reconnaissance. Conical targets, as typical structural or indicator targets, often require rapid and accurate positioning and identification in infrared thermal imaging applications to support subsequent tracking, guidance, or status assessment tasks.

[0003] Existing infrared thermal imaging target detection methods typically suffer from the following shortcomings: First, infrared images generally have relatively low resolution, significant noise interference, weak target edge features, and insufficient contrast, leading to unstable generation of target candidate regions and a tendency for missed detections and false detections. Second, when the background thermal radiation distribution of the scene is complex or the target scale varies greatly, traditional threshold- or rule-based detection methods are sensitive to parameters and lack generalization ability. Third, deep learning-based detection methods are prone to generating duplicate candidate boxes or low-confidence candidate results under conditions of weak infrared features; without an effective post-processing mechanism, the stability and usability of the detection results are affected.

[0004] Therefore, existing cone-shaped infrared thermal imaging detection methods have the problem of being unable to stably generate candidate regions and output accurate detection results under conditions of low contrast, low signal-to-noise ratio, and complex backgrounds. Summary of the Invention

[0005] This invention provides a cone-shaped infrared thermal imaging target recognition method based on an improved YOLOv5s, which solves the problem that existing cone-shaped infrared thermal imaging detection methods cannot stably generate candidate regions and output accurate detection results under conditions of low contrast, low signal-to-noise ratio and complex background.

[0006] In a first aspect, the present invention provides a cone-shaped infrared thermal imaging target identification method based on an improved YOLOv5s, the method comprising the following steps: Acquire infrared thermal imaging image data of the object to be detected; The infrared thermal imaging image data to be detected is input into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target. Post-processing is performed on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

[0007] Optionally, acquiring the infrared thermal imaging image data to be detected includes: Acquire infrared image frames or infrared video streams captured by an infrared thermal imaging sensor; Frame selection or timestamp alignment is performed on the infrared image frames or the infrared video stream to determine the infrared thermal imaging image data to be detected.

[0008] Optionally, before inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target, the method further includes: Obtain the infrared thermal imaging detection model to be trained; The feature extraction module in the infrared thermal imaging detection model to be trained is replaced with an improved feature extraction module to obtain the improved infrared thermal imaging detection model to be trained. The algorithm is optimized by using preset bounding box parameters to generate preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained. Based on the preset frame parameters, the improved infrared thermal imaging detection model to be trained is trained, and when the training converges, the preset infrared thermal imaging detection model is obtained.

[0009] Optionally, the step of generating preset bounding box parameters for training the improved infrared thermal imaging detection model by optimizing the generation algorithm using preset bounding box parameters includes: Obtain the width and height data of each cone-shaped target annotation box in the training dataset, and construct a set of sample box widths and heights; Use the width and height parameters of the preset anchor point set as the initial cluster centers; Calculate the IoU distance between each sample box in the sample box width and height set and the anchor point corresponding to the initial cluster center, and assign the sample box to the cluster corresponding to the cluster center with the smallest IoU distance; The cluster centers are updated based on the statistical characteristics of the width and height of the sample frames within each cluster, and it is determined whether the preset convergence conditions are met. When the preset convergence condition is met, the width and height parameters of the anchor point corresponding to the updated cluster center are determined as the preset frame parameters.

[0010] Optionally, the step of training the improved infrared thermal imaging detection model to be trained based on the preset bounding box parameters, and obtaining the preset infrared thermal imaging detection model upon training convergence, includes: Obtain the training dataset and the corresponding cone-shaped target bounding box data; The preset box parameters are used as the prediction prior parameters of the improved infrared thermal imaging detection model to be trained, and the prediction layer of the improved infrared thermal imaging detection model to be trained is configured. The training dataset is input into the improved infrared thermal imaging detection model to be trained to obtain the detection target; Based on the target bounding box data of the cone sleeve and the predicted prior parameters, the detected target is evaluated to determine the evaluation data of the detected target, which includes precision, recall and mean precision. Based on the evaluation data and the cone-shaped target bounding box data, the parameters of the improved infrared thermal imaging detection model to be trained are iteratively updated. When the evaluation data meets the preset evaluation convergence requirements, the preset infrared thermal imaging detection model is obtained.

[0011] Optionally, the step of inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target includes: The infrared thermal imaging image data to be detected is input into the backbone network of the preset infrared thermal imaging detection model to extract multi-level feature maps. The multi-level feature map is input into the neck network of the preset infrared thermal imaging detection model for cross-scale feature fusion processing to obtain a fused feature map; The fused feature map is input into the prediction layer of the preset infrared thermal imaging detection model to determine the prediction information of the cone-shaped target. The prediction information includes the location parameters, confidence parameters, and category parameters of the candidate region. Based on the location parameters, confidence parameters, and category parameters of the candidate regions, the target candidate regions for the cone sleeve are determined.

[0012] Optionally, the post-processing of the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target includes: Obtain the confidence parameter and position parameter corresponding to the target candidate region of the conical sleeve; Based on the confidence parameter, the target candidate region of the conical sleeve is subjected to confidence screening, and candidate regions with confidence scores lower than the preset confidence threshold are removed to obtain a candidate region retention set; Calculate the overlap between each candidate region in the candidate region retention set, and perform non-maximum suppression processing based on the overlap to suppress duplicate candidate regions with an overlap greater than a preset overlap threshold, thereby obtaining the target retention set; The target retention set is determined as the cone sleeve target detection result.

[0013] Secondly, the present invention also provides a cone-shaped infrared thermal imaging target recognition device based on an improved YOLOv5s, the cone-shaped infrared thermal imaging target recognition device based on the improved YOLOv5s comprising: The first acquisition module is used to acquire the infrared thermal imaging image data to be detected; The first reasoning module is used to input the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for reasoning processing to obtain the cone-shaped target candidate region. The first determining module is used to perform post-processing on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

[0014] Thirdly, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5s provided by the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5s provided by the invention.

[0016] This invention acquires infrared thermal imaging image data to be detected; inputs the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain candidate regions for cone-shaped targets; performs post-processing on the candidate regions to determine the detection result of the cone-shaped target, wherein the post-processing includes confidence filtering and non-maximum suppression. Through the above method steps, rapid identification and stable output of cone-shaped targets can be achieved in infrared thermal imaging scenarios, reducing the risk of missed detections and false detections, and is suitable for infrared target detection applications with high requirements for real-time performance and accuracy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a cone-shaped infrared thermal imaging target recognition method based on an improved YOLOv5s according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a feature extraction component framework provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a feature extraction module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of an improved feature extraction module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a cone-shaped infrared thermal imaging target recognition device based on an improved YOLOv5S provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, Figure 1 This is a flowchart of a cone-shaped infrared thermal imaging target recognition method based on an improved YOLOv5s according to an embodiment of the present invention. The method includes the following steps: 101. Acquire the infrared thermal imaging image data to be detected.

[0021] In this embodiment of the invention, the above-mentioned cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5s can be applied to the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s. The cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s has functions such as infrared thermal imaging recognition data processing, infrared thermal imaging recognition data transmission and reception, and infrared thermal imaging recognition data memory storage. It can be built based on a server or server cluster. The server or server cluster can be an electronic device with infrared thermal imaging recognition data processing capabilities.

[0022] The infrared thermal imaging image data to be detected can be understood as infrared image information acquired by an infrared thermal imaging sensor and meeting the input specifications of the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s. It is used to characterize the difference in thermal radiation intensity between the cone-shaped target and the background, and serves as direct input data for subsequent inference processing.

[0023] The direct input data mentioned above can refer to a single frame of infrared image or a sequence of consecutive frames in an infrared video stream.

[0024] Specifically, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s obtains raw image frames / video streams by accessing an infrared thermal imaging camera, and optionally performs frame selection or timestamp alignment, such as extracting 1 frame every N frames of the video stream, or aligning multiple inputs according to the same timestamp. At the same time, the image size is scaled to the model input resolution and normalized to form the infrared thermal imaging image data to be detected.

[0025] 102. Input the infrared thermal imaging image data to be detected into the preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target.

[0026] In this embodiment of the invention, the aforementioned preset infrared thermal imaging detection model can be understood as the detection model that has been trained and solidified offline before the deployment of the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s. Its network structure is based on YOLOv5s and adapted for the cone-shaped infrared thermal imaging scenario, so that the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s only needs to load the model and perform forward inference during the running phase to output candidate regions.

[0027] Specifically, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can read the model weight file from the memory and load it into the processor / accelerator during the initialization phase, and directly call the preset model to perform inference on the image to be detected. It is understood that the aforementioned preset infrared thermal imaging detection model does not need to be retrained for each detection, thus avoiding the impact of training overhead on real-time performance and ensuring the availability of engineering deployment.

[0028] In one possible embodiment, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s performs forward calculations on the input infrared image and outputs prediction information under the condition that the parameters of the preset infrared thermal imaging detection model are fixed, thereby transforming the pixel-domain thermal imaging image into structured target prediction information.

[0029] Specifically, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s first extracts multi-level feature maps from the backbone network, then performs cross-scale feature fusion using the neck network, and finally outputs the predicted information of the cone-shaped target from the prediction layer. For example, in an infrared image frame containing a cone, the inference processing outputs multiple candidate boxes and their corresponding confidence scores, which are then used by subsequent post-processing to further filter out redundant and low-confidence candidates, thereby maintaining high recall even under complex thermal background conditions.

[0030] The aforementioned candidate region for the cone-shaped target can refer to a set of regions containing the cone-shaped target. It is generally represented by bounding boxes and carries information such as positional and confidence parameters, serving as input for post-processing. Specifically, when the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s outputs candidate regions, multiple overlapping candidate boxes may appear for the same cone-shaped target. For example, different scale feature layers may simultaneously provide detection boxes, or low-confidence pseudo-candidate boxes may appear due to background hotspot interference.

[0031] 103. Perform post-processing on the candidate regions of the cone-shaped target to determine the detection results of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

[0032] In this embodiment of the invention, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s performs confidence filtering and overlap redundancy removal on the candidate region set of cone-shaped targets to obtain the target detection result. Specifically, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s performs at least confidence filtering and non-maximum suppression on the candidate regions: first, low-confidence candidates are eliminated based on the confidence threshold, and then duplicate candidate boxes are suppressed based on the overlap threshold. Through this post-processing link, duplicate alarm boxes and false alarms caused by infrared image noise and background thermal interference can be avoided, making the final output more stable and facilitating subsequent tracking, alarming or control.

[0033] The aforementioned cone-shaped target detection results can refer to the final recognition result set output by the platform after post-processing. It is used to provide structured target information to upper-layer applications, including but not limited to the bounding box position parameters and corresponding confidence parameters of the cone-shaped target. If necessary, it can also output derived information such as the number of targets and the coordinates of the center point.

[0034] Specifically, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s completes confidence screening and non-maximum suppression, obtains the target retention set, and determines it as the cone-shaped target detection result, for example, outputting "target box (x1,y1,x2,y2), confidence p". By outputting the target set after redundancy removal, the interference of multi-frame jitter on the upper-level control logic can be reduced, thereby improving the overall stability and real-time response quality of the system.

[0035] In one possible embodiment, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can filter the candidate region set based on the confidence parameters corresponding to the candidate regions, eliminate low-confidence candidates to reduce the false detection rate, and reduce the computational load of subsequent non-maximum suppression.

[0036] In another possible embodiment, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can suppress redundant candidate boxes in the candidate retention set according to the overlap relationship between candidate regions, thereby avoiding duplicate box output caused by the same cone-shaped target being predicted multiple times.

[0037] In this embodiment of the invention, infrared thermal imaging image data to be detected is acquired; the infrared thermal imaging image data to be detected is input into a preset infrared thermal imaging detection model for inference processing to obtain candidate regions for the cone-shaped target; post-processing is performed on the candidate regions to determine the detection result of the cone-shaped target, including confidence filtering and non-maximum suppression. Through the above method steps, rapid identification and stable output of cone-shaped targets can be achieved in infrared thermal imaging scenarios, reducing the risk of missed detections and false detections, and is suitable for infrared target detection application scenarios with high requirements for real-time performance and accuracy.

[0038] Optionally, in the step of acquiring the infrared thermal imaging image data to be detected, infrared image frames or infrared video streams collected by the infrared thermal imaging sensor can also be acquired; frame selection or timestamp alignment can be performed on the infrared image frames or infrared video streams to determine the infrared thermal imaging image data to be detected.

[0039] In this embodiment of the invention, the above-mentioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can acquire infrared image frames or infrared video streams collected by infrared thermal imaging sensors, and perform frame selection or timestamp alignment on the infrared image frames or infrared video streams to determine the infrared thermal imaging image data to be detected for inference.

[0040] Specifically, when the infrared thermal imaging sensor outputs in video mode, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can select target frames as input from the infrared video stream according to a preset sampling strategy. For example, it can extract one frame every N frames, or prioritize extracting the frame at the moment when the scene change exceeds a threshold, so as to balance real-time performance and computational load. When the infrared thermal imaging sensor needs to be aligned with other acquisition sources (such as pose sensors or visible light cameras), the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can add a timestamp to each frame of infrared image, and select the infrared image frame closest to the target moment within a preset time window according to the timestamp, thereby completing the timestamp alignment and outputting the infrared thermal imaging image data to be detected.

[0041] For example, in the scenario where the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s receives infrared video streams at 25fps, N=5 can be set for frame selection, reducing the inference input frequency to 5fps to reduce computing power consumption; or when it is necessary to align with the 1Hz status data of the external control system, a timestamp alignment strategy can be used to select the closest infrared frame as the image data to be detected at each whole second.

[0042] By using the above methods and steps, we can reduce the invalid inference overhead caused by redundant frame inputs and reduce candidate region drift caused by inter-frame temporal inconsistencies, thereby improving the stability of cone-shaped target detection results and the real-time performance of the system.

[0043] Optionally, before inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target, the steps also include: obtaining an infrared thermal imaging detection model to be trained; replacing the feature extraction module in the infrared thermal imaging detection model to be trained with an improved feature extraction module to obtain an improved infrared thermal imaging detection model to be trained; generating preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained through a preset bounding box parameter optimization generation algorithm; and training the improved infrared thermal imaging detection model to be trained based on the preset bounding box parameters, obtaining the preset infrared thermal imaging detection model when the training converges.

[0044] In this embodiment of the invention, the infrared thermal imaging detection model to be trained may refer to the basic detection network selected by the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s during the offline training phase. Its parameters have not yet been adapted and optimized for the cone-shaped infrared thermal imaging data, so it needs to be iteratively updated in conjunction with the training data before it can be deployed and used. In one possible embodiment, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can directly use the target detection network based on YOLOv5s as the initial model, and use the cone-shaped labeled data to specifically learn the model in subsequent training, thereby shortening the convergence time and ensuring training stability.

[0045] The aforementioned feature extraction module can refer to the network part in the model to be trained that converts the input infrared thermal imaging image into a multi-level feature map. It is usually located in the backbone network and its associated feature enhancement structure. Its output feature map will be used by the neck network for cross-scale fusion and finally used by the prediction layer for classification and bounding box regression.

[0046] The aforementioned improved feature extraction module refers to the feature extraction structure that is replaced by the improved YOLOv5s-based cone infrared thermal imaging target recognition platform in the model to be trained, and is enhanced for the cone infrared thermal imaging scene. Its purpose is to improve the separability of cone target features under low infrared contrast and low signal-to-noise ratio conditions and suppress background thermal noise interference.

[0047] Specifically, the improved feature extraction module can enhance the perception of the overall structure of the cone by expanding the convolutional receptive field, and enhance the local significant response of the cone and suppress the interference of flat background or hot spot areas through multi-level feature aggregation / fusion paths, so that subsequent prediction layers can more easily generate candidate regions with high confidence and more stable localization.

[0048] Furthermore, it can be achieved through methods such as Figure 2 The schematic diagram of the feature extraction component framework shown below. Figure 3 The schematic diagram of the feature extraction module shown is as follows: Figure 4 The improved feature extraction module is illustrated in the schematic diagram shown below. Figure 2 The C3 module is a core component of the YOLO object detection model series, mainly used to enhance feature extraction capabilities, improve model efficiency, and increase robustness. The CBS module consists of three core operations chained together: convolution, batch normalization, and activation function. The Add module performs weighted fusion of feature maps at different scales. The Concat module concatenates feature maps by increasing the number of channels.

[0049] Figure 3 The diagram in the middle shows the structure of the feature extraction module C3. Figure 4 This is a schematic diagram of the improved feature extraction module C3. In this embodiment, the receptive field of the model during feature extraction is expanded by adjusting the size of the convolution kernel in the CBS module. This improvement stems from the characteristics of umbrella-shaped targets in infrared thermal imaging—although their outline features are visible in the overall image, the details of the target become blurred when magnified locally due to uneven thermal radiation distribution. Therefore, a larger convolution kernel is needed to capture a wider range of contextual information.

[0050] To address the issue of deep coupling between thermal radiation characteristics and background noise, a new CBS submodule is added to the C3 module. This submodule enhances the overall feature extraction capability for weak targets through multi-level feature aggregation, thereby suppressing noise interference.

[0051] The aforementioned improved infrared thermal imaging detection model to be trained can refer to the detection network obtained by replacing the original feature extraction module in the infrared thermal imaging detection model to be trained with an improved feature extraction module in the aforementioned cone-shaped infrared thermal imaging target recognition platform based on improved YOLOv5s. Its network structure has been structurally modified for infrared cone-shaped scenes, but the model parameters are still in a state of optimization. It needs to be iteratively trained with the training dataset to form the final deployable preset infrared thermal imaging detection model.

[0052] The aforementioned preset bounding box parameter optimization generation algorithm can refer to the algorithm flow used by the improved YOLOv5s-based cone infrared thermal imaging target recognition platform to generate anchor prior bounding box parameters that better match the scale distribution of the cone target before training. Its input is the width and height data of the cone target annotation boxes in the training dataset, and the output is a set of preset bounding box parameters used to configure the prediction layer.

[0053] For example, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can construct a set of sample frame width and height, use the width and height of the preset anchor point set as the initial cluster center, perform cluster assignment and center update based on the IoU distance between the sample frame and the cluster center, and output the updated cluster center as the preset frame parameter when the preset convergence condition is met.

[0054] The aforementioned preset bounding box parameters can be understood as a set of prior bounding box width and height parameters obtained by the preset bounding box parameter optimization generation algorithm, i.e., the width and height configuration of the Anchor set. The aforementioned cone-shaped infrared thermal imaging target recognition platform based on improved YOLOv5s uses these as predictive prior parameters and configures them into the prediction layer of the improved infrared thermal imaging detection model to be trained. This is used to establish a matching relationship with the labeled bounding boxes during training to determine the allocation of training samples and to provide a scale reference for bounding box regression. By adopting preset bounding box parameters that better fit the size distribution of the cone-shaped target, the training instability and localization deviation caused by "excessive difference between the prior bounding box and the real target" can be reduced, thereby improving detection accuracy, recall rate, and the stability of the final output in infrared weak feature scenes.

[0055] Through the above-described methods and steps, the improved YOLOv5s-based conical infrared thermal imaging target recognition platform first performs targeted modifications to the feature extraction module of the basic detection network in the offline stage. It then combines a pre-defined bounding box parameter optimization generation algorithm based on the width and height distribution of the training bounding boxes to obtain prior parameters that better match the scale of the conical target. These prior parameters are then used to assist in training the bounding boxes and achieve model convergence. In the online stage, the pre-defined model is used for inference, and confidence filtering and non-maximum suppression are used to ensure the credibility and redundancy removal of candidate results. This achieves the goal of more stably generating candidate regions and outputting consistent conical target detection results under conditions of low infrared contrast, low signal-to-noise ratio, and complex thermal backgrounds, reducing false positives and false negatives, and minimizing duplicate bounding boxes and box jitter.

[0056] Optionally, the step of generating preset bounding box parameters for training the improved infrared thermal imaging detection model through the preset bounding box parameter optimization generation algorithm further includes: obtaining the width and height data of each cone-shaped target annotation box in the training dataset to construct a sample bounding box width and height set; using the width and height parameters of the preset anchor point set as the initial cluster centers; calculating the IoU distance between each sample bounding box in the sample bounding box width and height set and the corresponding anchor point of the initial cluster center, and assigning the sample bounding box to the cluster corresponding to the cluster center with the smallest IoU distance; updating the cluster centers based on the width and height statistical characteristics of the sample bounding boxes in each cluster, and determining whether the preset convergence condition is met; when the preset convergence condition is met, determining the width and height parameters of the updated anchor point corresponding to the cluster center as the preset bounding box parameters.

[0057] In this embodiment of the invention, the training dataset can refer to a set of infrared thermal imaging samples used for offline training and parameter optimization of the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s. This set includes, but is not limited to, several infrared thermal imaging images or video frames and their corresponding cone-shaped target annotation information. These samples are used to support the preset bounding box parameter optimization generation algorithm to extract the target scale distribution and are used for subsequent model training convergence. The training dataset can also cover cone-shaped targets under different distances, angles, and background thermal interference conditions, such as hot spots, reflections, and equipment heating elements.

[0058] The aforementioned cone-shaped target annotation box can be understood as the ground truth information of the target location obtained by manual or semi-automatic annotation for each cone-shaped target in the training dataset. It can generally be represented in the form of a rectangular bounding box, which includes the coordinates and width and height parameters of the bounding box. It is used to characterize the actual occupied area of ​​the cone-shaped target in the image. For example, when the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s performs preset box parameter optimization, it can directly extract the width and height data from the annotation box as the sample scale input, so that the width and height of the preset anchor point are more consistent with the real scale distribution of the cone-shaped target, reducing the positive and negative sample matching bias during the training phase.

[0059] The aforementioned set of sample bounding box width and height can refer to the set of width w and height h data extracted from all the cone-shaped target bounding boxes in the training dataset. Each element corresponds to a (w,h) pair of a cone-shaped target instance and is used as the clustering sample input for the preset bounding box parameter optimization generation algorithm. For example, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can traverse each bounding box in the training dataset, normalize its width and height to the model input scale (such as 640×640), and then write it into the set, thereby ensuring that samples of different resolutions have a consistent scale reference during clustering, and thus improving the stability of the generated preset bounding box parameters.

[0060] The aforementioned set of preset anchor points can refer to a set of anchor point width and height parameters used for initializing clustering or as a priori reference, i.e., several pairs of (w,h) parameters, used to represent the prior box scale of the prediction layer of the target detection model. In this embodiment, the width and height of 9 preset anchor points on a general dataset can be selected as initial values, or a set of empirical anchor point width and height parameters can be given as a set of preset anchor points based on historical cone data, so as to quickly converge to the preset box parameters that fit the cone target scale distribution in subsequent clustering updates.

[0061] Generally, the default set of preset anchor points in Yolov5 can be... - [10,13, 16,30, 33,23]# P3 / 8 - [30,61, 62,45, 59,119]# P4 / 16 - [116,90, 156,198, 373,326]# P5 / 32 There are a total of 9 groups of 3: small-scale, medium-scale, and large-scale; each group consists of two numbers. For example: (10, 13) Width (w) = 10; Height (h) = 13; The default set of anchor points in Yolov5 is generated from the COCO dataset, which differs significantly from the scale of the targets in our data. This generally leads to reduced training efficiency, decreased detection accuracy, and weakened generalization ability.

[0062] Therefore, this embodiment introduces IoU distance to cluster a preset set of anchor points that is more suitable for the training dataset used in this embodiment, thereby improving the convergence speed of the model during training.

[0063] The above IoU distance can be obtained by the following formula: The IoU distance metric is calculated as follows:

[0064] The IoU is calculated as follows:

[0065] in The width and height of the target bounding box. These are the width and height values ​​of the final generated preset bounding boxes. This method dynamically assigns the bounding boxes to the nearest cluster by calculating the IoU (Intersection over Union) distance between the width and height of each target sample box and the centers of nine preset anchor points. The preset bounding box parameters generated based on this clustering result have a size ratio that matches the target box distribution of the cone-shaped data, enabling the model to converge quickly during training and improving the detection accuracy of cone-shaped targets.

[0066] The aforementioned initial cluster center can refer to each group (w,h) in the preset anchor point set as the initial center point of the clustering algorithm when executing the preset box parameter optimization generation algorithm, which is used to cluster the sample box width and height set; its function is to provide an initial scale reference for clustering, so that the subsequent iteration of "sample allocation - center update" can converge faster around the target scale distribution of the cone, avoiding the slow convergence or local optimum problem caused by completely random initialization.

[0067] The aforementioned width and height statistical features can refer to the statistical measures of the width and height data of sample frames within the same cluster. These features are used to update the cluster centers so that the updated anchor point width and height can represent the typical scale of the samples in that cluster. For example, the mean / median of the width of all samples in the cluster can be taken, and the mean / median of the height can be taken, or a representative value that maximizes the average IoU within the cluster can be used as the width and height statistical features, thereby obtaining cluster center parameters that better represent the target scale distribution of the cluster.

[0068] The aforementioned preset convergence condition can be understood as a threshold or rule for determining whether clustering iteration can be stopped. In one possible embodiment, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on improved YOLOv5s determines whether the condition is met after each update to determine whether to output the final preset bounding box parameters. For example, it determines whether the width and height changes of each cluster center in two consecutive iterations are less than a preset threshold, or whether the average IoU improvement within the cluster is less than a preset threshold. When the above conditions are met, the width and height parameters of the anchor points corresponding to the updated cluster centers are determined as preset bounding box parameters and output. These parameters are used to configure the prediction layer for subsequent model training, thereby controlling computational overhead and improving overall training efficiency while ensuring generation quality.

[0069] Optionally, the step of training the improved infrared thermal imaging detection model to be trained based on preset bounding box parameters, and obtaining the preset infrared thermal imaging detection model upon training convergence, further includes: acquiring the training dataset and the cone-shaped target bounding box data corresponding to the training dataset; using the preset bounding box parameters as the prediction prior parameters of the improved infrared thermal imaging detection model to be trained, and configuring the prediction layer of the improved infrared thermal imaging detection model to be trained; inputting the training dataset into the improved infrared thermal imaging detection model to be trained to obtain the detection target; evaluating the detection target based on the cone-shaped target bounding box data and the prediction prior parameters, and determining the evaluation data of the detection target, including precision, recall, and mean precision; iteratively updating the parameters of the improved infrared thermal imaging detection model to be trained based on the evaluation data and the cone-shaped target bounding box data, and obtaining the preset infrared thermal imaging detection model when the evaluation data meets the preset evaluation convergence requirements.

[0070] In this embodiment of the invention, the aforementioned prediction prior parameters may refer to the prior configuration parameters formed by the improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition platform after configuring preset box parameters, such as the anchor point width and height set, into the prediction layer of the improved infrared thermal imaging detection model to be trained. These parameters are used to provide scale references for the candidate boxes output by the model and participate in the training sample matching process.

[0071] Specifically, during training, positive and negative samples can be determined and regression targets can be generated based on the matching relationship between the prior bounding box and the cone-shaped target bounding box. This makes the bounding box regression learning more stable and converges faster. In particular, it can reduce the localization deviation caused by the large difference between the prior bounding box and the real bounding box in scenarios where the scale of the cone-shaped target changes greatly or the infrared weak feature is weak.

[0072] The aforementioned detection target can refer to the set of target prediction results output by the model prediction layer, including but not limited to several candidate boxes and their corresponding category predictions and confidence information, which are used to compare with the cone-shaped target annotation data to complete training evaluation and parameter updates. For example, for an infrared image containing a cone, the model may output multiple candidate boxes, each of which has positional parameters (coordinates or center point + width and height), confidence parameters, and category parameters.

[0073] In one possible embodiment, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can determine the detection performance of the model in the current training round based on the correspondence between the cone-shaped target bounding box data and the detected target. This includes, but is not limited to, matching the detected target with the bounding box (e.g., determining whether it is a hit based on the IoU threshold), and statistically analyzing the correct detection, incorrect detection, and missed detection, in order to obtain evaluation results for guiding training updates and convergence judgments.

[0074] The aforementioned evaluation data can refer to a set of quantitative indicators obtained after the evaluation process is completed. These indicators are used to reflect the model's detection performance on the cone-shaped target and serve as a basis for judging training convergence. They include, but are not limited to, precision, recall, and mean precision. It can be understood that the aforementioned evaluation data are calculated based on the statistical results of the validation set or training set after each or several rounds of training, thereby forming a traceable basis for the trend of model performance changes.

[0075] Furthermore, the training results of the model can be evaluated based on the above evaluation data, thus constructing a multi-dimensional evaluation framework for target recognition and localization capabilities.

[0076] The calculation of precision and recall relies on the confusion matrix theory. This matrix compares the ground truth labels with the predicted labels to classify the prediction results into four typical cases: when both the ground truth label and the predicted result are positive, it is recorded as a true positive (TP); when the ground truth label is positive but the model misclassifies it as negative, it is recorded as a false negative (FN); when the ground truth label is negative but the model misclassifies it as positive, it is recorded as a false positive (FP); and when both the ground truth label and the predicted result are negative, it is recorded as a true negative (TN).

[0077] Precision and recall are calculated as follows:

[0078]

[0079] The mAP is calculated as follows:

[0080] Where N is the total number of categories, AP is the average precision, and AP is the area under the precision-recall (PR) curve calculated using different IoU thresholds (e.g., 0.5).

[0081] The aforementioned precision is used to characterize the proportion of actual cone-shaped targets among those detected as cone-shaped targets by the improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition platform, thus measuring the false detection control capability. Recall is used to characterize the proportion of cone-shaped targets in the bounding box successfully detected by the improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition platform, thus measuring the false negative control capability. Mean accuracy (mAP) is used to comprehensively reflect the detection accuracy performance under different confidence thresholds and measure the overall detection capability. The improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition platform can statistically analyze TP / FP / FN results and calculate the aforementioned indicators during the evaluation phase (e.g., by calculating the area under the precision-recall curve to obtain AP, and averaging by category to obtain mAP), thereby evaluating the model's comprehensive performance under complex infrared backgrounds and weak target conditions.

[0082] The aforementioned preset evaluation convergence requirement can refer to the evaluation criteria used to determine whether the training has converged and can stop iterative updates. It can consist of one or more evaluation data thresholds and their stability conditions.

[0083] For example, the average precision can be set to reach a preset threshold and the improvement in M ​​consecutive evaluations is less than a preset incremental threshold, or the precision and recall can reach a preset lower limit and the fluctuation is less than a preset fluctuation range in several consecutive rounds. When the evaluation data meets the preset evaluation convergence requirements, the above-mentioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s is determined to have completed training and outputs the preset infrared thermal imaging detection model.

[0084] Optionally, the step of inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the hoop target further includes: inputting the infrared thermal imaging image data to be detected into the backbone network of the preset infrared thermal imaging detection model to extract multi-level feature maps; inputting the multi-level feature maps into the neck network of the preset infrared thermal imaging detection model for cross-scale feature fusion processing to obtain a fused feature map; inputting the fused feature map into the prediction layer of the preset infrared thermal imaging detection model to determine the prediction information of the hoop target, the prediction information including the position parameters, confidence parameters, and category parameters of the candidate region; and determining the candidate region of the hoop target based on the position parameters, confidence parameters, and category parameters of the candidate region.

[0085] In this embodiment of the invention, the above-mentioned multi-level feature map can be understood as a multi-scale feature representation output by the backbone network at different network depths, which is used to simultaneously characterize the detailed texture / edge information and high-level semantic contour information of the cone-shaped target.

[0086] Specifically, in the backbone network, the feature maps output by the shallow layer are usually high resolution and can preserve the edge and local temperature change features of the cone-shaped target, making them suitable for supporting the localization of small targets or weak edges. On the other hand, the feature maps output by the deep layer have lower resolution but stronger semantics and can characterize the overall shape of the cone-shaped target and the information distinguishing it from the background, making them suitable for supporting robust recognition in complex backgrounds. Therefore, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s extracts multi-level feature maps so that the preset infrared thermal imaging detection model can identify local significant responses under weak infrared contrast and suppress background thermal noise by using high-level semantics.

[0087] The aforementioned fused feature map can refer to a unified feature representation obtained through cross-scale feature fusion processing. Its function is to complement the information of features at different levels and resolutions, so that the response of the cone-shaped target at different scales is enhanced and aligned in the same fusion space.

[0088] Specifically, the neck network can upsample high-level semantic features and concatenate or weightedly fuse them with low-level detail features. At the same time, it can downsample low-level features and pass them on to high-level features to form a two-way information flow from top to bottom and from bottom to top. This allows the fused feature map to contain both detailed localization clues of the cone-shaped target and semantic suppression information for complex thermal backgrounds. Therefore, the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s can reduce the risk of missed detections caused by relying solely on single-scale features, such as small cone-shaped targets being submerged or falsely detected in low-resolution deep features, or the risk of misjudging background hot spots by relying solely on shallow edges.

[0089] The aforementioned prediction information may refer to the set of structured prediction results output after the prediction layer, including but not limited to the location parameters, confidence parameters, and category parameters of the candidate regions.

[0090] In simple terms, the aforementioned positional parameters can be represented as bounding box coordinates or center point coordinates and width and height. The aforementioned confidence parameter can be used to characterize the degree of confidence that the candidate region is a trap target. The aforementioned category parameter can be used to characterize the category to which the candidate region belongs, such as the trap category. The trap infrared thermal imaging target recognition platform based on the improved YOLOv5s can generate a set of trap target candidate regions based on this prediction information. For example, it can use multiple bounding boxes and their confidence scores output by the prediction layer as candidate regions and enter them into the post-processing stage.

[0091] Optionally, the step of performing post-processing on the candidate regions of the cone-shaped target and determining the detection result of the cone-shaped target further includes obtaining the confidence parameter and position parameter corresponding to the candidate regions of the cone-shaped target; performing confidence screening on the candidate regions of the cone-shaped target based on the confidence parameter, removing candidate regions with confidence scores lower than a preset confidence threshold, and obtaining a candidate region retention set; calculating the overlap between each candidate region in the candidate region retention set, and performing non-maximum suppression processing based on the overlap to suppress duplicate candidate regions with overlap scores greater than a preset overlap threshold, and obtaining a target retention set; and determining the target retention set as the cone-shaped target detection result.

[0092] In this embodiment of the invention, after obtaining the candidate regions of the cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s, the platform performs validity filtering on the candidate regions according to the confidence parameters corresponding to each candidate region. Specifically, in the post-processing stage, the confidence parameters of each candidate box are read and compared with a preset confidence threshold. Candidate boxes below the threshold are directly deleted, while candidate boxes above the threshold are added to the candidate region retention set. This allows subsequent non-maximum suppression to mainly act on more reliable candidate boxes, thereby improving the stability of the final output.

[0093] The aforementioned pre-set confidence threshold can refer to the confidence judgment threshold used to distinguish between "credible candidate regions" and "low-credible candidate regions". Generally speaking, the aforementioned pre-set confidence threshold can be configured according to the statistical results of the training set / validation set, the trade-off between false positives and false negatives, or the requirements of the scenario. For example, in scenarios with strong background thermal noise and high false positive risk, the threshold can be set to 0.6~0.8 to enhance filtering, while in scenarios requiring high recall, it can be set to 0.3~0.5 to reduce false negatives.

[0094] The aforementioned candidate region retention set can be the set of candidate regions retained after performing confidence filtering on the target candidate region of the cone sleeve. Its elements are candidate boxes whose confidence scores meet the preset confidence threshold conditions and their corresponding position parameters and confidence parameters.

[0095] The aforementioned overlap refers to the degree of spatial overlap between any two candidate boxes in the candidate region retention set. It is generally represented by the Intersection over Union (IoU), with a higher value indicating that the two candidate boxes are more likely to correspond to the same target. For example, the IoU can be obtained by calculating the ratio of the intersection area to the union area of ​​two candidate boxes, which can be used to determine whether there is significant duplicate prediction between candidate boxes.

[0096] In one possible embodiment, the aforementioned cone-shaped infrared thermal imaging target recognition platform based on the improved YOLOv5s performs redundancy removal on duplicate candidate boxes within the candidate region retention set based on the overlap between candidate boxes. Specifically, the candidate boxes can be sorted by confidence level, and the candidate box with the highest confidence level can be selected as the retention box. The overlap between the candidate box and the remaining candidate boxes can be calculated. If the overlap between a candidate box and the retention box exceeds a preset overlap threshold, it is determined to be a duplicate candidate and suppressed. This process is repeated until the traversal is complete, thereby outputting a more stable set of target boxes.

[0097] The aforementioned preset overlap threshold can refer to the overlap determination threshold used for non-maximum suppression processing. Generally, it can be configured according to the target size, positioning accuracy requirements and scene interference level. For example, it can be set to 0.4~0.6 to balance the deduplication effect and multi-target separation capability. If the threshold is too low, it may incorrectly suppress the candidate boxes of adjacent targets. If the threshold is too high, it may retain too many duplicate boxes, resulting in unstable output.

[0098] The target retention set obtained by the above method can be the deredundant cone-shaped target box and its corresponding confidence and category information, which can be directly used as the output of the cone-shaped target detection result. For example, when the same cone-shaped target is predicted multiple times, non-maximum suppression will merge multiple overlapping candidate boxes into a high-confidence target box, thereby reducing duplicate boxes, reducing box jitter and improving the stability of downstream tracking / control modules.

[0099] like Figure 5 As shown, this embodiment of the invention also provides a cone-shaped infrared thermal imaging target recognition device 500 based on an improved YOLOv5s, which includes: The first acquisition module 501 is used to acquire infrared thermal imaging image data to be detected; The first reasoning module 502 is used to input the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for reasoning processing to obtain the cone-shaped target candidate region. The first determining module 503 is used to perform post-processing on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

[0100] Optionally, the first acquisition module 501 mentioned above includes: The first acquisition submodule is used to acquire infrared image frames or infrared video streams collected by the infrared thermal imaging sensor; The second acquisition submodule is used to select frames or align timestamps on the infrared image frames or the infrared video stream to determine the infrared thermal imaging image data to be detected.

[0101] Optionally, the above-mentioned device further includes: The first training module is used to obtain the infrared thermal imaging detection model to be trained; The second training module is used to replace the feature extraction module in the infrared thermal imaging detection model to be trained with an improved feature extraction module to obtain the improved infrared thermal imaging detection model to be trained. The third training module is used to optimize the generation algorithm through preset bounding box parameters to generate preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained. The fourth training module is used to train the improved infrared thermal imaging detection model to be trained based on the preset box parameters, and obtain the preset infrared thermal imaging detection model when the training converges.

[0102] Optionally, the third training module mentioned above includes: The first training submodule is used to obtain the width and height data of each cone-shaped target annotation box in the training dataset and construct a set of sample box widths and heights. The second training submodule is used to take the width and height parameters of the preset anchor point set as the initial cluster centers; The third training submodule is used to calculate the IoU distance between each sample box in the sample box width and height set and the anchor point corresponding to the initial cluster center, and to assign the sample box to the cluster corresponding to the cluster center with the smallest IoU distance; The fourth training submodule is used to update the cluster centers based on the statistical features of the width and height of the sample boxes within each cluster, and to determine whether the preset convergence conditions are met. When the preset convergence condition is met, the width and height parameters of the anchor point corresponding to the updated cluster center are determined as the preset frame parameters.

[0103] Optionally, the fourth training module mentioned above includes: The fifth training submodule is used to acquire the training dataset and the cone-shaped target annotation box data corresponding to the training dataset; The sixth training submodule is used to use the preset box parameters as the prediction prior parameters of the improved infrared thermal imaging detection model to be trained, and to configure the prediction layer of the improved infrared thermal imaging detection model to be trained. The seventh training submodule is used to input the training dataset into the improved infrared thermal imaging detection model to be trained, so as to obtain the detection target; The eighth training submodule is used to evaluate the detected target based on the cone-shaped target bounding box data and the predicted prior parameters, and to determine the evaluation data of the detected target, which includes precision, recall and mean precision. The ninth training submodule is used to iteratively update the parameters of the improved infrared thermal imaging detection model to be trained based on the evaluation data and the cone-shaped target annotation data. When the evaluation data meets the preset evaluation convergence requirements, a preset infrared thermal imaging detection model is obtained.

[0104] Optionally, the first inference module 502 mentioned above includes: The first inference submodule is used to input the infrared thermal imaging image data to be detected into the backbone network of the preset infrared thermal imaging detection model and extract multi-level feature maps. The second inference submodule is used to input the multi-level feature map into the neck network of the preset infrared thermal imaging detection model for cross-scale feature fusion processing to obtain a fused feature map. The third inference submodule is used to input the fused feature map into the prediction layer of the preset infrared thermal imaging detection model to determine the prediction information of the cone-shaped target. The prediction information includes the location parameters, confidence parameters, and category parameters of the candidate region. The fourth inference submodule is used to determine the target candidate region of the cone sleeve based on the position parameters, confidence parameters, and category parameters of the candidate region.

[0105] Optionally, the first determining module 503 mentioned above includes: The first determining submodule is used to obtain the confidence parameter and position parameter corresponding to the candidate region of the conical sleeve target; The second determining submodule is used to perform confidence screening on the target candidate region of the conical sleeve based on the confidence parameter, and remove candidate regions with confidence scores lower than a preset confidence threshold to obtain a candidate region retention set; The third determining submodule is used to calculate the overlap between each candidate region in the candidate region retention set, and perform non-maximum suppression processing based on the overlap to suppress duplicate candidate regions with an overlap greater than a preset overlap threshold, thereby obtaining the target retention set. The fourth determination submodule is used to determine the target retention set as the cone sleeve target detection result.

[0106] like Figure 6 As shown, this embodiment of the invention also provides an electronic device 600, including a processor, which can execute any of the above-mentioned cone-shaped infrared thermal imaging target recognition methods based on improved YOLOv5s.

[0107] Specifically, it includes a processor 601 and a memory 602, as well as a computer program stored in the memory 602 and capable of running on the processor 601, which executes a cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5s, wherein: The processor 601 executes the calculator program stored in the memory 602, which is based on the improved YOLOv5s cone-shaped infrared thermal imaging target recognition method, and performs the following steps: Acquire infrared thermal imaging image data of the object to be detected; The infrared thermal imaging image data to be detected is input into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target. Post-processing is performed on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

[0108] Optionally, the processor 601 performs the acquisition of the infrared thermal imaging image data to be detected, including: Acquire infrared image frames or infrared video streams captured by an infrared thermal imaging sensor; Frame selection or timestamp alignment is performed on the infrared image frames or the infrared video stream to determine the infrared thermal imaging image data to be detected.

[0109] Optionally, before the processor 601 executes the step of inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target, the method further includes: Obtain the infrared thermal imaging detection model to be trained; The feature extraction module in the infrared thermal imaging detection model to be trained is replaced with an improved feature extraction module to obtain the improved infrared thermal imaging detection model to be trained. The algorithm is optimized by using preset bounding box parameters to generate preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained. Based on the preset frame parameters, the improved infrared thermal imaging detection model to be trained is trained, and when the training converges, the preset infrared thermal imaging detection model is obtained.

[0110] Optionally, the processor 601 executes the optimization generation algorithm based on preset bounding box parameters to generate preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained, including: Obtain the width and height data of each cone-shaped target annotation box in the training dataset, and construct a set of sample box widths and heights; Use the width and height parameters of the preset anchor point set as the initial cluster centers; Calculate the IoU distance between each sample box in the sample box width and height set and the anchor point corresponding to the initial cluster center, and assign the sample box to the cluster corresponding to the cluster center with the smallest IoU distance; The cluster centers are updated based on the statistical characteristics of the width and height of the sample frames within each cluster, and it is determined whether the preset convergence conditions are met. When the preset convergence condition is met, the width and height parameters of the anchor point corresponding to the updated cluster center are determined as the preset frame parameters.

[0111] Optionally, the processor 601 executes the training of the improved infrared thermal imaging detection model to be trained based on the preset bounding box parameters, and obtains the preset infrared thermal imaging detection model upon training convergence, including: Obtain the training dataset and the corresponding cone-shaped target bounding box data; The preset box parameters are used as the prediction prior parameters of the improved infrared thermal imaging detection model to be trained, and the prediction layer of the improved infrared thermal imaging detection model to be trained is configured. The training dataset is input into the improved infrared thermal imaging detection model to be trained to obtain the detection target; Based on the target bounding box data of the cone sleeve and the predicted prior parameters, the detected target is evaluated to determine the evaluation data of the detected target, which includes precision, recall and mean precision. Based on the evaluation data and the cone-shaped target bounding box data, the parameters of the improved infrared thermal imaging detection model to be trained are iteratively updated. When the evaluation data meets the preset evaluation convergence requirements, the preset infrared thermal imaging detection model is obtained.

[0112] Optionally, the processor 601 performs the step of inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target, including: The infrared thermal imaging image data to be detected is input into the backbone network of the preset infrared thermal imaging detection model to extract multi-level feature maps. The multi-level feature map is input into the neck network of the preset infrared thermal imaging detection model for cross-scale feature fusion processing to obtain a fused feature map; The fused feature map is input into the prediction layer of the preset infrared thermal imaging detection model to determine the prediction information of the cone-shaped target. The prediction information includes the location parameters, confidence parameters, and category parameters of the candidate region. Based on the location parameters, confidence parameters, and category parameters of the candidate regions, the target candidate regions for the cone sleeve are determined.

[0113] Optionally, the processor 601 performs the post-processing on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target, including: Obtain the confidence parameter and position parameter corresponding to the target candidate region of the conical sleeve; Based on the confidence parameter, the target candidate region of the conical sleeve is subjected to confidence screening, and candidate regions with confidence scores lower than the preset confidence threshold are removed to obtain a candidate region retention set; Calculate the overlap between each candidate region in the candidate region retention set, and perform non-maximum suppression processing based on the overlap to suppress duplicate candidate regions with an overlap greater than a preset overlap threshold, thereby obtaining the target retention set; The target retention set is determined as the cone sleeve target detection result.

[0114] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition method or the application-side improved YOLOv5s-based cone-shaped infrared thermal imaging target recognition method provided in this invention, and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0115] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be done by a computer program instructing related hardware, and can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0116] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A target recognition method based on improved YOLOv5s conical infrared thermal imaging, characterized in that, include: Acquire the infrared thermal imaging image data of the object to be detected; The infrared thermal imaging image data to be detected is input into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target. Post-processing is performed on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

2. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 1, characterized in that, The acquisition of the infrared thermal imaging image data to be detected includes: Acquire infrared image frames or infrared video streams captured by an infrared thermal imaging sensor; Frame selection or timestamp alignment is performed on the infrared image frames or the infrared video stream to determine the infrared thermal imaging image data to be detected.

3. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 1, characterized in that, Before inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target, the method further includes: Obtain the infrared thermal imaging detection model to be trained; The feature extraction module in the infrared thermal imaging detection model to be trained is replaced with an improved feature extraction module to obtain the improved infrared thermal imaging detection model to be trained. The algorithm is optimized by using preset bounding box parameters to generate preset bounding box parameters for training the improved infrared thermal imaging detection model to be trained. Based on the preset frame parameters, the improved infrared thermal imaging detection model to be trained is trained, and when the training converges, the preset infrared thermal imaging detection model is obtained.

4. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 3, characterized in that, The step of generating preset bounding box parameters through a preset bounding box parameter optimization algorithm to train the improved infrared thermal imaging detection model includes: Obtain the width and height data of each cone-shaped target annotation box in the training dataset, and construct a set of sample box widths and heights; Use the width and height parameters of the preset anchor point set as the initial cluster centers; Calculate the IoU distance between each sample box in the sample box width and height set and the anchor point corresponding to the initial cluster center, and assign the sample box to the cluster corresponding to the cluster center with the smallest IoU distance; The cluster centers are updated based on the statistical characteristics of the width and height of the sample frames within each cluster, and it is determined whether the preset convergence conditions are met. When the preset convergence condition is met, the width and height parameters of the anchor point corresponding to the updated cluster center are determined as the preset frame parameters.

5. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 3, characterized in that, The improved infrared thermal imaging detection model to be trained is trained based on the preset bounding box parameters. When the training converges, the preset infrared thermal imaging detection model is obtained, including: Obtain the training dataset and the corresponding cone-shaped target bounding box data; The preset box parameters are used as the prediction prior parameters of the improved infrared thermal imaging detection model to be trained, and the prediction layer of the improved infrared thermal imaging detection model to be trained is configured. The training dataset is input into the improved infrared thermal imaging detection model to be trained to obtain the detection target; Based on the target bounding box data of the cone sleeve and the predicted prior parameters, the detected target is evaluated to determine the evaluation data of the detected target, which includes precision, recall and mean precision. Based on the evaluation data and the cone-shaped target bounding box data, the parameters of the improved infrared thermal imaging detection model to be trained are iteratively updated. When the evaluation data meets the preset evaluation convergence requirements, the preset infrared thermal imaging detection model is obtained.

6. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 1, characterized in that, The step of inputting the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for inference processing to obtain the candidate region of the cone-shaped target includes: The infrared thermal imaging image data to be detected is input into the backbone network of the preset infrared thermal imaging detection model to extract multi-level feature maps. The multi-level feature map is input into the neck network of the preset infrared thermal imaging detection model for cross-scale feature fusion processing to obtain a fused feature map; The fused feature map is input into the prediction layer of the preset infrared thermal imaging detection model to determine the prediction information of the cone-shaped target. The prediction information includes the location parameters, confidence parameters, and category parameters of the candidate region. Based on the location parameters, confidence parameters, and category parameters of the candidate regions, the target candidate regions for the cone sleeve are determined.

7. The cone-shaped infrared thermal imaging target recognition method based on improved YOLOv5s as described in claim 6, characterized in that, The post-processing of the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target includes: Obtain the confidence parameter and position parameter corresponding to the target candidate region of the conical sleeve; Based on the confidence parameter, the target candidate region of the conical sleeve is subjected to confidence screening, and candidate regions with confidence scores lower than the preset confidence threshold are removed to obtain a candidate region retention set; Calculate the overlap between each candidate region in the candidate region retention set, and perform non-maximum suppression processing based on the overlap to suppress duplicate candidate regions with an overlap greater than a preset overlap threshold, thereby obtaining the target retention set; The target retention set is determined as the cone sleeve target detection result.

8. A cone-shaped infrared thermal imaging target recognition device based on an improved YOLOv5S, characterized in that, include: The first acquisition module is used to acquire the infrared thermal imaging image data to be detected; The first reasoning module is used to input the infrared thermal imaging image data to be detected into a preset infrared thermal imaging detection model for reasoning processing to obtain the cone-shaped target candidate region. The first determining module is used to perform post-processing on the candidate region of the cone-shaped target to determine the detection result of the cone-shaped target. The post-processing includes confidence screening and non-maximum suppression.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps in the cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5S as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps in the cone-shaped infrared thermal imaging target recognition method based on the improved YOLOv5S as described in any one of claims 1 to 7.

Images