Infrared image photovoltaic module hot spot defect identification method, device, equipment and medium

Abstract

The invention discloses an infrared image photovoltaic module hot spot defect identification method and device, equipment and a medium, and relates to the technical field of deep learning. The method comprises the following steps: acquiring a photovoltaic module infrared image data set with a photovoltaic module hot spot defect marking bounding box; constructing a target detection model, wherein the target detection model comprises an ESRGAN network, an improved Transform network and a detection head; performing iterative training on the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect identification model; and identifying the infrared image of the to-be-identified photovoltaic module through the hot spot defect identification model to obtain hot spot defect information on the infrared image of the to-be-identified photovoltaic module. In this way, the ESRGAN network and the improved Transform network are utilized to improve the recognition precision of the to-be-recognized photovoltaic module infrared image with low resolution.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method, device, equipment and medium for identifying hot spot defects of photovoltaic components using infrared images. Background Art

[0002] In the inspection of photovoltaic modules, the infrared images of photovoltaic modules collected by drones can more accurately identify hot spots, heating problems or local temperature anomalies, and discover hidden surface defects or problems. However, the resolution of infrared images is usually lower than that of visible light images, which affects the detection and positioning accuracy of small defects. Summary of the invention

[0003] In view of this, the purpose of the present invention is to overcome the deficiencies in the prior art and to provide a method, device, equipment and medium for identifying hot spot defects in photovoltaic components using infrared images, which are used to identify hot spot defect information on infrared images of photovoltaic components to be identified with lower resolution.

[0004] The present invention provides the following technical solutions:

[0005] In a first aspect, the present invention provides a method for identifying hot spot defects of photovoltaic modules using infrared images, comprising:

[0006] Obtain a photovoltaic module infrared image dataset with annotated bounding boxes of photovoltaic module hot spot defects;

[0007] Building a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head;

[0008] Iteratively training the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model;

[0009] The infrared image of the photovoltaic assembly to be identified is identified by using the hot spot defect identification model to obtain the hot spot defect information on the infrared image of the photovoltaic assembly to be identified.

[0010] In one embodiment, the target detection model is iteratively trained according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model, including:

[0011] Improving the resolution of the photovoltaic module infrared image dataset by using the ESRGAN network to obtain a high-resolution image dataset;

[0012] During the i-th round of training, feature extraction is performed based on the high-resolution image data set by using the improved Transformer network to obtain a target feature map;

[0013] The detection head performs prediction according to the target feature map to obtain a target prediction result;

[0014] Calculating the high-resolution image dataset and the target prediction result based on a loss function to obtain a loss value;

[0015] The model parameters of the target detection model of the i-th round are optimized according to the loss value of the i-th round to obtain the target detection model of the i+1-th round, until the iteration stop condition is met, the iterative training is stopped, and the target detection model after the iterative training is used as the hot spot defect recognition model.

[0016] In one embodiment, the target detection model further includes a backbone network, and the feature extraction is performed according to the high-resolution image data set by the improved Transformer network to obtain a target feature map, including:

[0017] Performing feature extraction on the high-resolution image dataset through the improved Transformer network to obtain a first multi-scale feature map;

[0018] Performing feature extraction on the high-resolution image data set through the backbone network to obtain a second multi-scale feature map;

[0019] Feature fusion is performed on the first multi-scale feature map and the second multi-scale feature map to obtain the target feature map.

[0020] In one embodiment, extracting features from the high-resolution image dataset using the improved Transformer network to obtain a first multi-scale feature map includes:

[0021] Using the improved Transformer network, the high-resolution image dataset is subjected to step-by-step feature extraction using a hierarchical structure to obtain image features at multiple stages;

[0022] Based on the sparse attention mechanism, feature enhancement is performed on the image features of each stage to obtain the first multi-scale feature map.

[0023] In one embodiment, the target detection model further includes a PAN path aggregation network, and the target prediction result is obtained by predicting according to the target feature map through the detection head, including:

[0024] Performing feature aggregation on the target feature graph through the PAN path aggregation network to obtain an aggregation result;

[0025] The detection head performs prediction according to the convergence result to obtain a target prediction result.

[0026] In one embodiment, the target detection model further includes a consistent dual allocation network, and the detection head performs prediction according to the convergence result to obtain a target prediction result, including:

[0027] Predicting the converged result by the detection head to obtain an initial prediction result;

[0028] The initial prediction result is processed by the consistent dual assignment network to obtain the target prediction result.

[0029] In one embodiment, the detection head includes a one-to-one detection head and a one-to-many detection head, the initial prediction result includes a first prediction result and a second prediction result, and the prediction of the converged result by the detection head to obtain the initial prediction result includes:

[0030] Predicting the convergence result by using the one-to-one detection head to obtain the first prediction result;

[0031] The convergence result is predicted by the one-to-many detection head to obtain the second prediction result.

[0032] In a second aspect, the present invention provides an infrared image photovoltaic module hot spot defect recognition device, comprising:

[0033] An acquisition module, used for acquiring a photovoltaic module infrared image dataset with annotated boundary boxes of photovoltaic module hot spot defects;

[0034] A construction module, used to construct a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head;

[0035] A training module, used for iteratively training the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model;

[0036] The recognition module is used to recognize the infrared image of the photovoltaic component to be recognized by using the hot spot defect recognition model to obtain the hot spot defect information on the infrared image of the photovoltaic component to be recognized.

[0037] In a third aspect, the present invention proposes a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the method for identifying hot spot defects of photovoltaic modules using infrared images as described in the first aspect is implemented.

[0038] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for identifying hot spot defects of photovoltaic modules using infrared images as described in the first aspect.

[0039] The method, device, equipment and medium for identifying hot spot defects of photovoltaic components in infrared images disclosed in the present invention obtain a data set of infrared images of photovoltaic components with annotated boundary boxes of hot spot defects of photovoltaic components; construct a target detection model, the target detection model includes an ESRGAN network, an improved Transformer network and a detection head; iteratively train the target detection model according to the infrared image data set of photovoltaic components to obtain a hot spot defect recognition model; identify the infrared image of the photovoltaic component to be identified through the hot spot defect recognition model to obtain the hot spot defect information on the infrared image of the photovoltaic component to be identified. In this way, the infrared image of the photovoltaic component to be identified with a lower resolution is processed using the ESRGAN network, so that the infrared image has clearer features; and feature extraction is performed in combination with the improved Transformer network to focus on the local features of the image, especially the prominent positions in the image, so as to improve the recognition accuracy of the infrared image of the photovoltaic component to be identified with a lower resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope of protection of the present invention. In each of the drawings, similar components are numbered similarly.

[0041] Figure 1 A schematic diagram of a process of the infrared image photovoltaic module hot spot defect recognition method proposed in this embodiment is shown;

[0042] Figure 2 Another schematic diagram of the process of the infrared image photovoltaic module hot spot defect recognition method proposed in this embodiment is shown;

[0043] Figure 3 Another schematic diagram of the process of the infrared image photovoltaic module hot spot defect recognition method proposed in this embodiment is shown;

[0044] Figure 4 Another schematic diagram of the process of the infrared image photovoltaic module hot spot defect recognition method proposed in this embodiment is shown;

[0045] Figure 5 A schematic structural diagram of the infrared image photovoltaic module hot spot defect recognition device proposed in this embodiment is shown.

[0046] Description of the accompanying drawings:

[0047] 500 - infrared image photovoltaic module hot spot defect recognition device; 501 - acquisition module; 502 - construction module; 503 - training module; 504 - recognition module. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0049] The components of the embodiments of the present invention generally described and shown in the drawings herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0050] Hereinafter, the terms "including", "having" and their cognates, which may be used in various embodiments of the present invention, are intended only to indicate specific features, numbers, steps, operations, elements, components or combinations of the foregoing items, and should not be understood as first excluding the existence of one or more other features, numbers, steps, operations, elements, components or combinations of the foregoing items or adding the possibility of one or more features, numbers, steps, operations, elements, components or combinations of the foregoing items.

[0051] Furthermore, the terms “first”, “second”, “third”, etc. are merely used for distinguishing descriptions and are not to be understood as indicating or implying relative importance.

[0052] Unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meanings as those generally understood by those skilled in the art to which the various embodiments of the present invention belong. The terms (such as those defined in generally used dictionaries) will be interpreted as having the same meanings as the contextual meanings in the relevant technical field and will not be interpreted as having idealized meanings or overly formal meanings unless clearly defined in the various embodiments of the present invention.

[0053] Example 1

[0054] The disclosed embodiment provides a method for identifying hot spot defects of photovoltaic components using infrared images, which is used to identify hot spot defect information on an infrared image of a photovoltaic component to be identified with a low resolution.

[0055] See also Figure 1 , a method for identifying hot spot defects of photovoltaic modules using infrared images includes steps S101 to S104, and each step is described in detail below.

[0056] Step S101 , obtaining a photovoltaic module infrared image dataset with annotated boundary boxes of photovoltaic module hot spot defects.

[0057] In this embodiment, a photovoltaic module infrared image dataset is obtained, which has a bounding box for hot spot defects of photovoltaic modules. Hot spot defects can be other types such as local hot spots and hot spots. In addition, the images in the dataset may be taken under different lighting conditions, shooting angles and distances, thereby increasing the diversity and complexity of the data.

[0058] Step S102, constructing a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head.

[0059] In this embodiment, a target detection model is constructed, and the network structure of the target detection model may include an Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) network, an improved Transformer network, and a detection head.

[0060] Among them, the target detection model can be a Trans-YOLOv10 network model.

[0061] The original YOLOv10 achieved significant speed and accuracy advantages over the previous YOLO algorithm in the COCO test of target detection algorithms. However, in infrared image target detection applications, since infrared images capture infrared radiation emitted by objects and reflect the thermal radiation characteristics of objects, which is different from visible light images, the YOLOv10 algorithm does not extract obvious features for infrared images, local features are not prominent, and the targets are small, which increases the difficulty of target recognition. The identification of hot spot faults in photovoltaic modules of drone infrared images belongs to the local identification of small targets in infrared images, which poses a greater challenge to the application of YOLOv10. It is difficult to achieve high detection accuracy by directly applying the YOLOv10 algorithm.

[0062] Combining YOLOv10 with the improved Transformer can greatly improve accuracy while taking into account the inference speed. This is a hybrid application of the latest target detection algorithm and the self-attention mechanism of the improved Transformer network. It is effective in identifying the hot spot fault location of photovoltaic modules in drone infrared images.

[0063] It should be noted that the ESRGAN network is used to generate high-resolution images, retaining more details and texture information; the improved Transformer network is used to extract image features. The improved Transformer network adopts a hierarchical architecture to extract multi-scale features step by step, and uses a sparse attention mechanism for fusion between layers, allowing the model to focus on important parts of the input data, which has a significant improvement in extracting prominent local features of the image; the detection head is used to predict results based on the extracted features.

[0064] Step S103, iteratively training the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model.

[0065] In this embodiment, the target detection model is iteratively trained based on the photovoltaic module infrared image dataset to update the model parameters of the target detection model until a preset number of iterations is reached or the loss value converges below a certain threshold, the iteration is stopped, and the iterated target detection model is used as the hot spot defect recognition model.

[0066] Step S104, identifying the infrared image of the photovoltaic assembly to be identified by using the hot spot defect recognition model to obtain hot spot defect information on the infrared image of the photovoltaic assembly to be identified.

[0067] In this embodiment, the resolution of the infrared image of the photovoltaic module to be identified is improved by training the ESRGAN network in the mature hot spot defect recognition model, and an image with clear texture is obtained as the input image of the subsequent model, which makes feature extraction easier and the distinction between defective parts and normal parts higher; and the improved Transformer network adopts a hierarchical architecture to extract multi-scale features step by step, and uses a sparse attention mechanism between layers for fusion and enhancement, allowing the model to focus on important parts of the input data, which has a significant improvement effect on extracting prominent local features of the image; finally, the detection head is used to make predictions based on the extracted features to more accurately obtain the detection result hot spot defect information such as the location of the hot spot fault on the infrared image of the photovoltaic module to be identified with a lower resolution, meeting the requirements of immediacy and high precision.

[0068] See also Figure 2 In a specific embodiment, step S103 includes steps S1031 to S1035, and each step is described in detail below.

[0069] Step S1031: improving the resolution of the photovoltaic module infrared image dataset through the ESRGAN network to obtain a high-resolution image dataset.

[0070] In this embodiment, the infrared image dataset of the photovoltaic module is input into the ESRGAN network to generate a high-resolution image dataset.

[0071] Step S1032: During the i-th round of training, feature extraction is performed based on the high-resolution image data set by using the improved Transformer network to obtain a target feature map.

[0072] In this embodiment, during the i-th round of training, the improved Transformer network performs feature extraction and feature fusion based on the high-resolution image data set to obtain a target feature map for subsequent model prediction. Since the input of the improved Transformer network is a high-resolution image data set, the feature details extracted by the improved Transformer network are clearer.

[0073] See also Figure 3 In a specific embodiment, the target detection model also includes a backbone network, and step S1032 includes steps S301 to S303. Each step is described in detail below.

[0074] Step S301: extract features from the high-resolution image dataset using the improved Transformer network to obtain a first multi-scale feature map.

[0075] In this embodiment, the first multi-scale feature map is obtained by performing feature extraction and feature enhancement operations based on the high-resolution image data set through an improved Transformer network.

[0076] In a specific embodiment, step S301 includes: using the improved Transformer network to perform step-by-step feature extraction on the high-resolution image dataset using a hierarchical structure to obtain multiple stage image features; and performing feature enhancement on each stage image feature based on a sparse attention mechanism to obtain the first multi-scale feature map.

[0077] In this example, see Figure 4 , the images in the high-resolution image dataset are input as input images to the improved Transformer network, and the hierarchical structure is used to extract features step by step to obtain image features at multiple stages. Feature_maps trans = Transformer(x in ), where x in Represents the input image in the high-resolution image dataset, Feature_maps transRepresents the multi-stage image features extracted by the Transformer network, and Transformer represents the Transformer feature extraction network.

[0078] For each stage image feature extracted, the query, key and value are first calculated: Q = zW Q K=zW K V=zW V , where z represents the matrix after the image is divided into blocks, W Q , W K , W V is the learned weight matrix, Q, K, V are the query, key and value matrices respectively.

[0079] Next, calculate the attention weights: Among them, d k is the dimension of the key, which is used for scaling to avoid too concentrated attention distribution. Through this mechanism, the model can dynamically focus on different parts of the image and obtain different features.

[0080] Furthermore, the image features extracted by the sparse attention mechanism are assigned attention weights according to their characteristics and scaled proportionally. Among them, I(x,y) represents the weight value of each position (x,y) in the image, α represents the scaling factor, when I(x,y) is less than 0.5, α takes the value of 0.6, in other cases, α takes the value of 1.2, Represents the scaled weight value.

[0081] Finally, the enhanced first multi-scale feature map is obtained Feature_maps trans represents the Transformer feature extraction network, Represents the first multi-scale feature map after enhancement, α represents the scaling factor, when Feature_maps trans When (x,y) is less than 0.5, α is 0.6, and in other cases, α is 1.2.

[0082] Step S302: extract features from the high-resolution image data set through the backbone network to obtain a second multi-scale feature map.

[0083] In this embodiment, the cross-stage partial network (CSPNet) in the backbone network is used to extract features from the high-resolution image dataset to obtain the second multi-scale feature map. CSPNet is an enhanced network that can improve gradient flow transfer and reduce computational redundancy. main =conv_layers(x in ), x in Represents the input image in the high-resolution image dataset, Feature_maps main It represents the second multi-scale feature map extracted by the backbone network, and conv_layers represents the backbone feature extraction network.

[0084] Step S303: performing feature fusion on the first multi-scale feature map and the second multi-scale feature map to obtain the target feature map.

[0085] In this embodiment, the first multi-scale feature map and the second multi-scale feature map are fused to obtain a fused target feature map to achieve more accurate and efficient model performance. The backbone network is good at extracting local features of the image, while the Transformer network is good at capturing global dependencies. The combination of the two enables the model to achieve better performance when processing complex tasks. Among them, Feature fusion represents the fused target feature map, Represents the first multi-scale feature map after Transformer extracts feature enhancement, Feature_maps main Represents the second multi-scale feature map extracted by the backbone network.

[0086] Please see again Figure 2 , step S1033, predicting according to the target feature map through the detection head to obtain a target prediction result.

[0087] In this embodiment, prediction is performed based on the fused target feature map by detecting the head, and a target prediction result with a prediction annotation box is output.

[0088] In a specific embodiment, the target detection model also includes a PAN path aggregation network, and step S1033 includes: performing feature aggregation on the target feature map through the PAN path aggregation network to obtain an aggregation result; and performing prediction according to the aggregation result through the detection head to obtain a target prediction result.

[0089] In this embodiment, the target feature map is sent to the PAN path aggregation network for multi-path aggregation. In the PAN path aggregation network, the target feature map will undergo bottom-up path enhancement and top-down path enhancement to achieve cross-scale feature fusion.

[0090] After the aggregation process of the PAN network, the converged result is a series of feature maps that fuse information of different scales. These feature maps usually have the same resolution and number of channels and contain rich contextual information and target details.

[0091] Furthermore, the detection head makes predictions based on the converged results to obtain target prediction results, including the bounding box coordinates, category, and confidence of each target.

[0092] In a specific embodiment, the target detection model also includes a consistent dual allocation network, and the detection head performs prediction based on the convergence result to obtain a target prediction result, including: predicting the convergence result through the detection head to obtain an initial prediction result; processing the initial prediction result through the consistent dual allocation network to obtain the target prediction result.

[0093] The detection head includes a one-to-one detection head and a one-to-many detection head, the initial prediction result includes a first prediction result and a second prediction result, and the prediction of the convergence result by the detection head to obtain the initial prediction result includes: predicting the convergence result by the one-to-one detection head to obtain the first prediction result; predicting the convergence result by the one-to-many detection head to obtain the second prediction result.

[0094] In this embodiment, the Consistent Dual Assignment Network is a strategy for optimizing the prediction results of the target detection model. It combines one-to-many and one-to-one label assignment methods to improve the accuracy and efficiency of the model.

[0095] The converged results are predicted by a one-to-one detection head to obtain a first prediction result, so as to assign multiple prediction boxes to each true label to provide rich supervision signals; the converged results are predicted by a one-to-many detection head to obtain a second prediction result, so as to assign only one best prediction box to each true label to avoid the use of non-maximum suppression (NMS), thereby improving the inference speed.

[0096] The consistent dual assignment network adopts a consistent matching metric to evaluate the consistency between the predicted box and the true label. By combining the one-to-many and one-to-one label assignment methods, it ensures that the model obtains rich supervision signals during training.

[0097] After being processed by the consistent dual assignment network, the redundancy and overlap in the initial prediction results are optimized, and only one best prediction box is retained for each object. The final object prediction results include the precise bounding box coordinates and category probabilities of each object. result =F consistent (OneToOne result ,OneToMany result ), where OneToMany result Represents the first prediction result of a one-to-many detection head, OneToOne result represents the second prediction result of the one-to-one detection head, F consistent Represents a consistent dual assignment network, ConsistentMatch result Represents the target prediction result after processing by the consistent dual assignment network.

[0098] Step S1034, calculating the high-resolution image data set and the target prediction result based on the loss function to obtain a loss value.

[0099] In this embodiment, the loss function is a function that measures the difference between the model prediction result and the actual result. The loss function is used to calculate the loss value of the resolution image dataset and the target prediction result, which is used to guide the target detection model to optimize the model parameters.

[0100] Step S1035, optimize the model parameters of the target detection model of the i-th round according to the loss value of the i-th round, obtain the target detection model of the i+1-th round, until the iteration stop condition is met, stop the iterative training, and use the target detection model after the iterative training as the hot spot defect recognition model.

[0101] In this embodiment, according to the loss value of the i-th round, the model parameters of the target detection model of the i-th round are updated using the selected optimization algorithm, and the updated model parameters will be used for the next round of training, that is, the target detection model of the i+1-th round is obtained. Until the preset iteration stop condition is met, the iterative training of the target detection model is stopped, and the target detection model after the iterative training is stopped is used as the hot spot defect recognition model. At this time, the hot spot defect recognition model can accurately identify the hot spot defect information on the infrared image of the photovoltaic component to be identified with a lower resolution.

[0102] Common iteration stopping conditions include the preset number of iterations, the loss value no longer decreases significantly (i.e. convergence), and the model's performance on the validation set no longer improves.

[0103] The infrared image photovoltaic module hot spot defect recognition method proposed in this embodiment obtains a photovoltaic module infrared image dataset with a photovoltaic module hot spot defect annotated boundary box; constructs a target detection model, the target detection model includes an ESRGAN network, an improved Transformer network and a detection head; iteratively trains the target detection model according to the photovoltaic module infrared image dataset to obtain a hot spot defect recognition model; recognizes the infrared image of the photovoltaic module to be recognized by the hot spot defect recognition model to obtain the hot spot defect information on the infrared image of the photovoltaic module to be recognized. In this way, the ESRGAN network is used to process the infrared image of the photovoltaic module to be recognized with a lower resolution, so that the infrared image has clearer features; and the improved Transformer network is combined to perform feature extraction to focus on the local features of the image, especially the prominent positions in the image, so as to improve the recognition accuracy of the infrared image of the photovoltaic module to be recognized with a lower resolution.

[0104] Example 2

[0105] In addition, the present disclosure provides an infrared image photovoltaic module hot spot defect recognition device 500, see Figure 5 , the device comprises:

[0106] An acquisition module 501 is used to acquire a photovoltaic module infrared image dataset with a photovoltaic module hot spot defect annotated boundary box;

[0107] A construction module 502 is used to construct a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head;

[0108] A training module 503 is used to iteratively train the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model;

[0109] The identification module 504 is used to identify the infrared image of the photovoltaic component to be identified by using the hot spot defect identification model to obtain the hot spot defect information on the infrared image of the photovoltaic component to be identified.

[0110] Optionally, the training module 503 is also used to improve the resolution of the photovoltaic module infrared image dataset through the ESRGAN network to obtain a high-resolution image dataset; in the i-th round of training, perform feature extraction based on the high-resolution image dataset through the improved Transformer network to obtain a target feature map; perform prediction based on the target feature map through the detection head to obtain a target prediction result; calculate the high-resolution image dataset and the target prediction result based on the loss function to obtain a loss value; optimize the model parameters of the i-th round target detection model according to the i-th round loss value to obtain the i+1-th round target detection model, until the iteration stop condition is met, stop the iterative training, and use the iteratively trained target detection model as the hot spot defect recognition model.

[0111] Optionally, the target detection model also includes a backbone network, and the trained module 503 is also used to perform feature extraction on the high-resolution image dataset through the improved Transformer network to obtain a first multi-scale feature map; perform feature extraction on the high-resolution image dataset through the backbone network to obtain a second multi-scale feature map; and perform feature fusion on the first multi-scale feature map and the second multi-scale feature map to obtain the target feature map.

[0112] Optionally, the training module 503 is also used to perform step-by-step feature extraction on the high-resolution image dataset using a hierarchical structure through the improved Transformer network to obtain multiple stage image features; and perform feature enhancement on each stage image feature based on a sparse attention mechanism to obtain the first multi-scale feature map.

[0113] Optionally, the target detection model also includes a PAN path aggregation network, and the training module 503 is also used to perform feature aggregation on the target feature map through the PAN path aggregation network to obtain an aggregation result; and perform prediction according to the aggregation result through the detection head to obtain a target prediction result.

[0114] Optionally, the target detection model also includes a consistent dual allocation network, and the training module 503 is also used to predict the convergence result through the detection head to obtain an initial prediction result; and process the initial prediction result through the consistent dual allocation network to obtain the target prediction result.

[0115] Optionally, the detection head includes a one-to-one detection head and a one-to-many detection head, and the training module 503 is further used to predict the convergence result through the one-to-one detection head to obtain the first prediction result; and predict the convergence result through the one-to-many detection head to obtain the second prediction result.

[0116] The device provided in the embodiment of the present disclosure can execute the steps of the infrared image photovoltaic module hot spot defect identification method provided in Example 1, which will not be described again to avoid repetition.

[0117] The infrared image photovoltaic module hot spot defect recognition device proposed in this embodiment obtains a photovoltaic module infrared image dataset with a photovoltaic module hot spot defect annotated boundary box; constructs a target detection model, the target detection model includes an ESRGAN network, an improved Transformer network and a detection head; iteratively trains the target detection model according to the photovoltaic module infrared image dataset to obtain a hot spot defect recognition model; recognizes the infrared image of the photovoltaic module to be recognized through the hot spot defect recognition model to obtain the hot spot defect information on the infrared image of the photovoltaic module to be recognized. In this way, the ESRGAN network is used to process the infrared image of the photovoltaic module to be recognized with a lower resolution, so that the infrared image has clearer features; and the improved Transformer network is combined to perform feature extraction to focus on the local features of the image, especially the prominent positions in the image, so as to improve the recognition accuracy of the infrared image of the photovoltaic module to be recognized with a lower resolution.

[0118] Example 3

[0119] In addition, an embodiment of the present disclosure provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the infrared image photovoltaic component hot spot defect identification method described in Example 1 is implemented.

[0120] The device provided in the embodiment of the present disclosure can execute the steps of the infrared image photovoltaic module hot spot defect identification method provided in Example 1, which will not be described again to avoid repetition.

[0121] Example 4

[0122] The disclosed embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the infrared image photovoltaic module hot spot defect identification method described in Embodiment 1 is implemented.

[0123] In this embodiment, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0124] The computer-readable storage medium provided in this embodiment can implement the infrared image photovoltaic module hot spot defect identification method provided in Example 1, and will not be described again here to avoid repetition.

[0125] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.

[0126] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0127] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for identifying hot spot defects of photovoltaic modules using infrared images, characterized in that: include: Obtain a photovoltaic module infrared image dataset with annotated bounding boxes of photovoltaic module hot spot defects; Building a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head; Iteratively training the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model; The infrared image of the photovoltaic assembly to be identified is identified by using the hot spot defect identification model to obtain the hot spot defect information on the infrared image of the photovoltaic assembly to be identified.

2. The infrared image photovoltaic module hot spot defect recognition method according to claim 1 is characterized in that: The iterative training of the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model includes: Improving the resolution of the photovoltaic module infrared image dataset by using the ESRGAN network to obtain a high-resolution image dataset; During the i-th round of training, feature extraction is performed based on the high-resolution image data set by using the improved Transformer network to obtain a target feature map; The detection head performs prediction according to the target feature map to obtain a target prediction result; Calculating the high-resolution image dataset and the target prediction result based on a loss function to obtain a loss value; The model parameters of the target detection model of the i-th round are optimized according to the loss value of the i-th round to obtain the target detection model of the i+1-th round, until the iteration stop condition is met, the iterative training is stopped, and the target detection model after the iterative training is used as the hot spot defect recognition model.

3. The infrared image photovoltaic module hot spot defect recognition method according to claim 2 is characterized in that: The target detection model also includes a backbone network, and the improved Transformer network is used to extract features according to the high-resolution image data set to obtain a target feature map, including: Performing feature extraction on the high-resolution image dataset through the improved Transformer network to obtain a first multi-scale feature map; Performing feature extraction on the high-resolution image data set through the backbone network to obtain a second multi-scale feature map; Feature fusion is performed on the first multi-scale feature map and the second multi-scale feature map to obtain the target feature map.

4. The method for identifying hot spot defects of photovoltaic modules using infrared images according to claim 3, characterized in that: The step of extracting features from the high-resolution image dataset using the improved Transformer network to obtain a first multi-scale feature map includes: Using the improved Transformer network, the high-resolution image dataset is subjected to step-by-step feature extraction using a hierarchical structure to obtain image features at multiple stages; Based on the sparse attention mechanism, feature enhancement is performed on the image features of each stage to obtain the first multi-scale feature map.

5. The infrared image photovoltaic module hot spot defect identification method according to claim 2 is characterized in that: The target detection model also includes a PAN path aggregation network, and the detection head performs prediction according to the target feature map to obtain a target prediction result, including: Performing feature aggregation on the target feature graph through the PAN path aggregation network to obtain an aggregation result; The detection head performs prediction according to the convergence result to obtain a target prediction result.

6. The infrared image photovoltaic module hot spot defect recognition method according to claim 5, characterized in that: The target detection model further includes a consistent dual allocation network, wherein the detection head performs prediction according to the convergence result to obtain a target prediction result, including: Predicting the converged result by the detection head to obtain an initial prediction result; The initial prediction result is processed by the consistent dual assignment network to obtain the target prediction result.

7. The method for identifying hot spot defects of photovoltaic modules using infrared images according to claim 6, characterized in that: The detection head includes a one-to-one detection head and a one-to-many detection head, the initial prediction result includes a first prediction result and a second prediction result, and the prediction of the converged result by the detection head to obtain the initial prediction result includes: Predicting the convergence result by using the one-to-one detection head to obtain the first prediction result; The convergence result is predicted by the one-to-many detection head to obtain the second prediction result.

8. An infrared image photovoltaic module hot spot defect recognition device, characterized in that: include: An acquisition module, used for acquiring a photovoltaic module infrared image dataset with annotated boundary boxes of photovoltaic module hot spot defects; A construction module, used to construct a target detection model, wherein the target detection model includes an ESRGAN network, an improved Transformer network, and a detection head; A training module, used for iteratively training the target detection model according to the photovoltaic module infrared image data set to obtain a hot spot defect recognition model; The recognition module is used to recognize the infrared image of the photovoltaic component to be recognized by using the hot spot defect recognition model to obtain the hot spot defect information on the infrared image of the photovoltaic component to be recognized.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the infrared image photovoltaic component hot spot defect recognition method as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the infrared image photovoltaic component hot spot defect recognition method as described in any one of claims 1 to 7.

Images