Low-illumination wide-field image change detection method and device based on depth discrimination

Abstract

This invention discloses a method and apparatus for detecting changes in low-light wide-field-of-view images based on depth discrimination. The method includes: creating training and validation datasets on dual-temporal image data for training, focusing on object and noise change features, and training ON-DisNet; calculating a global difference image on test data and retrieving changed objects based on a preset wide energy threshold; normalizing and performing morphological fusion filtering on the retrieved changed objects; using the trained ON-DisNet to determine whether the retrieved changed objects are objects or noise, and obtaining local image pairs from the input dual-temporal test images based on the changes, orientation, and scale information of real objects; extracting local difference features from the local image pairs using WeibDO and creating local difference images; applying LoG-Graph to perform binary segmentation on the local difference images to generate local change maps, and synthesizing them into a global change map. The apparatus includes a processor and a memory.

Description

Technical Field

[0001] This invention relates to the field of image change detection, and more particularly to a method and apparatus for detecting changes in low-light wide-field images based on depth discrimination. Background Technology

[0002] Under low-light conditions, wide-field-of-view video images are characterized by complex backgrounds, low contrast, and limited dynamic range. The structural and contour features of moving objects tend to degrade significantly, making it difficult for existing object detection algorithms trained under normal lighting conditions to capture and identify effective object features, resulting in a high false negative rate. It is worth noting that while the original features of objects may be distorted due to insufficient light, as long as they are not completely occluded or in absolute darkness, the presence of an object will always cause its background to exhibit significant differences in texture, structure, edges, and color compared to other object-free backgrounds.

[0003] Existing change detection algorithms can be mainly divided into three categories: pixel-based methods, object-based methods, and deep learning-based methods. Pixel-based methods are highly sensitive to scene dynamics (such as swaying leaves and fluctuations in lighting) and are not suitable for handling globally changing scenes with significant interference. Object-based methods involve object recognition and classification, requiring more computational resources and being highly sensitive to parameter selection. Deep learning-based methods rely on large amounts of labeled data for training and have insufficient generalization ability across different types of datasets. As mentioned earlier, models trained under normal illumination conditions cannot effectively detect moving objects in low-light wide-field-of-view images.

[0004] Currently, there is a lack of publicly available datasets for low-light wide-field-of-view video images, especially those related to change detection. Change detection involves image segmentation, ultimately resulting in a classification of each pixel in the image. Segmentation datasets require accurate pixel-level labeling of the boundaries of each changing object in the image, a complex and time-consuming process. Due to the scarcity of relevant segmentation datasets, training change detection models for low-light wide-field-of-view video images is challenging. In contrast, classification datasets only require overall classification or labeling of the image containing the object to be detected to indicate its presence, making them relatively easy to create. However, low-light wide-field-of-view video images are characterized by low contrast and complex backgrounds, resulting in blurred structures and contours of moving objects in the scene. Therefore, directly creating classification datasets for low-light wide-field-of-view video images is also quite difficult.

[0005] The main challenge in change detection of wide-field-of-view video images under low-light conditions is noise interference. The imaging process of digital images is primarily accomplished by sensors and other related electronic components, involving complex optical and electronic engineering. Under low-light conditions, imaging sensors need to increase sensitivity or extend exposure time to collect more light and improve image brightness, but this also leads to a significant increase in thermal noise and random photon noise, resulting in a noticeable deterioration in image quality. Therefore, change detection algorithms that directly target the global input image are susceptible to interference from a large amount of random noise or illumination fluctuations in unchanged background areas. Furthermore, moving objects occupy a relatively small percentage of pixels in wide-field-of-view images, and directly detecting them across the entire image would result in significant computational redundancy. Summary of the Invention

[0006] This invention provides a method and apparatus for detecting changes in low-light wide-field-of-view images based on depth discrimination. The invention trains a discriminative model on a self-created dataset of object and noise change features to distinguish changes caused by objects and noise, thereby achieving change detection based on the local change region where the real object is located. Details are described below: A method for detecting changes in low-light wide-field-of-view images based on depth discrimination, the method comprising:

[0007] Training and validation datasets on object and noise variation features are created on bi-temporal image data used for training, and ON-DisNet is trained on them.

[0008] A global difference image is calculated on the test data, and the changed objects are retrieved according to a preset wide energy threshold; the retrieved changed objects are normalized and morphologically fused and filtered.

[0009] The trained ON-DisNet is used to determine whether the retrieved changing object is an object or noise, and local image pairs are obtained from the input bi-temporal test image based on the changes, orientation and scale information of the real object.

[0010] WeibDO is used to extract local difference features from local image pairs and create local difference images; LoG-Graph is applied to perform binary segmentation on the local difference images to generate local change maps, and these are then combined into a global change map.

[0011] Specifically, calculating the global difference image on the test data involves:

[0012]

[0013] in, and These are the mean values ​​of the input biphase sequence images T1 and T2, respectively.

[0014] The normalization and morphological fusion filtering of the retrieved changed objects are as follows:

[0015] Morphological filtering is performed on the normalized candidate objects using structuring elements E1, E2, and E3 to fuse and amplify the variation features:

[0016]

[0017] Where, ω l Represents the weighting coefficient, L represents the number of objects to be merged, and I represents the weighting coefficient. input Indicates the input image, I f This indicates the fusion result.

[0018] The step of using WeibDO to extract local difference features from local image pairs is as follows:

[0019]

[0020] Where k and λ are the shape parameter and scale parameter, respectively, and I represents the extreme pixel ratio. The calculation principle is as follows:

[0021]

[0022] Where I1 and I2 represent dual-phase input sequences, and min and max represent the minimum and maximum values ​​of the pixels to be retained for comparison, respectively.

[0023] The method further includes: using the LoG operator to construct edge weights for the graph structure, the principle of which is:

[0024]

[0025] Where G represents the Gaussian function, and LoG represents the LoG operator. σ is the given standard deviation, set to 3.

[0026] Part Two: A low-light wide-field-of-view image change detection device based on depth discrimination, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to execute any of the methods described in Part One.

[0027] Part Three, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any of the first parts.

[0028] The beneficial effects of the technical solution provided by this invention are:

[0029] 1. This invention does not directly perform change detection on the global difference map of the original dual-temporal images. Instead, it learns the feature differences between objects and noise to distinguish them, and then detects the local region where the real target is located to avoid the interference of noise. Furthermore, it proposes and verifies for the first time a deep learning-based object-and-noise-oriented discriminative network (ON-DisNet) instead of relying on subjective threshold selection methods for discrimination.

[0030] 2. Before classifying targets and noise, this invention introduces a normalization module to unify the energy distribution of objects with different intensity of change in a globally differentiated image. Then, it performs classification operations on objects and noise at a more differentiated feature level, such as texture and contour, instead of directly classifying potential changing objects. In fact, the performance of target detection algorithms based on object features drops sharply in low-light wide-field-of-view video images mainly because the feature patterns they learn do not match the feature patterns degraded under low-light conditions. The most obvious degradation phenomenon of object features in low-light wide-field-of-view video images is the decrease in contrast between the foreground and background and the blurring of object texture and contour caused by dynamic distribution compression. In addition, this invention also uses Multi-Directional Structuring Elements-Based Morphological Fusion Filtering (MDSE-MFF) to quickly eliminate texture interference features, thereby further improving the classifier's ability to distinguish between objects with weak intensity of change and noise with strong interference features.

[0031] 3. This method does not rely solely on the fusion and improvement of existing difference operators to achieve performance enhancement. Based on the research and analysis of the essence of difference feature extraction, it explores the Weibull distribution curve with excellent variation characteristics across disciplines and proposes for the first time a Weibull-based Difference Operator (WeibDO). This operator has excellent noise interference suppression and variation feature enhancement performance, and is very suitable for extracting difference features in local variation scenes with significant interference features. Furthermore, it proposes a Laplacian-of-Gaussian-based Graph-Cut Algorithm (LoG-Graph), which achieves good segmentation performance for the local difference features extracted by WeibullDO. Attached Figure Description

[0032] Figure 1 A flowchart illustrating the change detection process based on an object- and noise-oriented discrimination model designed for this invention.

[0033] Figure 2 Schematic diagrams of structural elements in different forms;

[0034] Figure 3 The input is a dual-temporal video image;

[0035] Figure 4 To generate a change map by directly performing global change detection on the input bi-temporal image using WeibDO and LoG-Graph;

[0036] Figure 5 The resulting graph shows the changes after incorporating the deep learning discriminant model designed in this invention.

[0037] Figure 6 The resulting change diagram is generated after incorporating the discrimination model and optimization module designed in this invention;

[0038] Figure 7 The image shows the detection results of other comparison algorithms. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0040] Example 1

[0041] A change detection method for low-light wide-field-of-view video images, the method comprising the following steps:

[0042] Step 101: Create training and validation datasets on the bi-temporal image data used for training to feature changes in objects and noise, and train ON-DisNet.

[0043] Step 102: Calculate the Global Difference Image (GDI) on the test data and retrieve the changed objects according to the preset wide energy threshold;

[0044] Step 103: Normalize and perform MDSE-MFF operations on the retrieved changed objects;

[0045] Step 104: Use the trained ON-DisNet to determine whether the retrieved changed object is an object or noise, retain the change, orientation and scale information of the real object, and obtain local image pairs from the input dual-temporal test image based on this information;

[0046] Step 105: Use WeibDO to extract local difference features and create a Local Difference Image (LDI);

[0047] Step 106: Apply LoG-Graph to perform binary segmentation on LDI, generate Local Change Graph (LCI), and synthesize them into Global Change Graph (GCI).

[0048] In summary, the embodiments of the present invention train a discriminative model on a self-made dataset of change features of objects and noise to distinguish change information generated by objects and noise, thereby achieving change detection based on the local change region where the real object is located.

[0049] Example 2

[0050] The scheme in Example 1 will be further described below with reference to specific calculation formulas and accompanying drawings. See the description below for details:

[0051] Step 201: Before creating a dataset of features of changes in objects and noise or performing subsequent local change detection, it is necessary to first generate GDI and retrieve potential objects with changes.

[0052] In this embodiment of the invention, an improved LRO (ILRO) is used to directly generate GDI based on the input dual-temporal image, as shown in formula (1).

[0053]

[0054] in, and These are the mean values ​​of the input biphase sequence images T1 and T2, respectively.

[0055] Step 202: Due to the lack of publicly available datasets specifically for distinguishing between objects and noise, both model training and validation were performed on a self-collected object and noise dataset. This dataset consists of 1200 samples, each 96×96 pixels, with objects and noise each comprising half. The object and noise samples were used to form the training and validation sets in a 9:1 ratio.

[0056] Step 203: ON-DisNet is designed based on the lightweight residual network ResNet18. The specific training parameters and optimization configuration settings are as follows:

[0057] The initial learning rate and training batch size were set to 0.0001 and 64, respectively, with a training period of 10 epochs. Considering the relatively small size of the training dataset, the adaptive mechanism of the Adam optimizer was utilized to minimize the need for manual hyperparameter tuning, reduce the risk of vanishing or exploding gradients, and achieve fast convergence. The Adam optimizer needs to calculate the loss function relative to the parameter φ. t The gradient of g is denoted as g. tThen, the biased first-order moment estimate and the biased second-order moment estimate are calculated respectively. Subsequently, the deviation of the moment estimate is adjusted according to formulas (2) and (3).

[0058]

[0059]

[0060] Where ξ1 and ξ2 represent the exponential decay rates. Finally, φ is updated according to formula (4). t .

[0061]

[0062] Here, Γ represents the learning rate, and ε represents a small scalar used to enhance numerical stability. Furthermore, a step learning rate scheduler is employed, halving the learning rate every 5 epochs. Simultaneously, the cross-entropy loss is calculated by integrating the log-soft maximum function and the negative log-likelihood (NLL) loss, and the difference between the predicted and actual distributions is quantified according to formula (5).

[0063]

[0064] in, Let y represent the predicted probability of the i-th class. y is a one-hot encoded binary vector where only the positions related to the actual class are marked as 1 and the rest are marked as 0, thus effectively focusing the computation of NLL loss on the true class.

[0065] Step 204: Compared with objects with higher intensity of change, objects with relatively weaker intensity of change have a narrower grayscale dynamic range, making them more likely to be misjudged as noise.

[0066] According to formula (6), the potential change object is normalized to effectively balance the dynamic range of the change object with different energy characteristics, so that the discrimination model can rely more on structural features such as texture and shape to distinguish objects and noise, and reduce the possibility of false negatives and false positives in the detection results.

[0067]

[0068] Among them, P min P represents the minimum value in the input image. max This represents the maximum value. α represents the normalization scale range, set to 255.

[0069] Under low illumination conditions, noise primarily manifests as patchy noise related to lighting variations, striped noise related to buildings, and random spot noise. The impact of this type of noise can be significantly reduced by strategically applying structural elements with a clear directionality. This invention defines three types of structural elements, E1, E2, and E3, as follows: Figure 4 As shown. Subsequently, morphological filtering is performed on the normalized candidate objects using structuring elements E1, E2, and E3 to selectively eliminate interfering features. Then, the changing features are fused and amplified according to formula (7).

[0070]

[0071] Where, ω l Represents the weighting coefficient, L represents the number of objects to be merged, and I represents the weighting coefficient. input Indicates the input image, I f This indicates the fusion result. This embodiment of the invention employs three different structural elements, with L = 3.

[0072] Step 205: Use WeibDO to extract the differential features of local regions of interest pairs and generate LDI. The calculation principle is shown in formula (8).

[0073]

[0074] Where k and λ are the shape parameter and scale parameter, respectively. I represents the extreme pixel ratio, and its calculation principle is shown in formula (9).

[0075]

[0076] Where I1 and I2 represent dual-phase input sequences, and min and max represent the minimum and maximum values ​​of the pixels to be retained for comparison, respectively.

[0077] To enhance the contrast between the changed and unchanged regions, it is necessary to suppress the relatively weak variations caused by noise in the unchanged regions (where the ratio of corresponding pixels in the two-phase sequence is close to 1), while further amplifying the significant difference features produced by objects in the changed regions (where the ratio of corresponding pixels in the two-phase sequence is close to 0). Therefore, k = 0.5 and λ = 0.5 are set, respectively. Before performing binary segmentation, morphological closing (CO) and median filtering (MF) are used to optimize the feature representation of LDI.

[0078] Step 205: Represent the LDI as a graph, where each pixel represents a node. Foreground and background segmentation can be achieved by minimizing the graph-based energy equation. The energy equation, as shown in Equation (10), quantifies the cost of graph segmentation.

[0079]

[0080] Here, the energy function E(L) represents the sum of D(L) and V(L). D(L) is the penalty term D on the data. p The summation describes the assignment of label L to each pixel p in the image set P. p The cost. On the other hand, V(L) is a smoothing term, representing the label pair (L) assigned to each pair of adjacent pixels (p, q) in the neighborhood set B. p L q Interactive potential energy V between ) p，q Summation of .

[0081] This method uses the LoG operator to construct the edge weights of the graph structure, and its principle is shown in formula (11).

[0082]

[0083] Where G represents the Gaussian function and LoG represents the LoG operator. σ is the given standard deviation, set to 3. The generated LoG kernel is used to perform a convolution operation on the input image to obtain the corresponding LoG response, which contains the edge and texture information of each pixel in the image.

[0084] Finally, the minimum cut of the graph structure is achieved by using the optimized maximum flow solution method

[14] , resulting in LCI, which is then synthesized into GCI.

[0085] Example 3

[0086] The schemes in Examples 1 and 2 will be further described below with specific examples:

[0087] First, the performance of WeibDO was compared with existing difference operators such as SO[1], LRO[2], MRO[3], NRO[4], and EPRO[5], and the tests were conducted on three low-light wide-field-of-view video images with different interference features. In order to eliminate the possibility that the adjustable parameters in the segmentation algorithm are biased towards a specific operator, the unmodified K-means algorithm was used to cluster the difference features extracted by different operators to generate change maps, so as to objectively evaluate the performance differences of different operators. Table 1 lists the average performance evaluation results of different operators, where MRO was evaluated under three window sizes (#3, #5, and #7). The experimental results show that WeibDO has the best overall performance, with a lower false alarm rate and a higher detection accuracy.

[0088] Table 1 Evaluation values ​​of different operators

[0089]

[0090] Then, the overall algorithm was compared with existing change detection algorithms such as PCAK[6], CDIK[7], NR_ELM[8], CWNN[9], ASEA

[10] , Zhu

[11] , NPSG

[12] , INLPG

[13] and Shi[5]. Table 3 lists the evaluation results of the ten change detection methods. CDIK performed well in S1 and S2, with KC values ​​of 79.94% and 75.95%, respectively. However, CDIK is very sensitive to scene changes, especially in the more challenging S3 scene, where the lighting is low and random noise increases, resulting in a significant drop in performance. PCAK and ASEA showed a similar pattern to CDIK, with KC values ​​of 67.42% and 68.02% in the S1 scene, respectively, but dropping to 1.95% and 0.31% in the S3 scene. Other methods, including CWNN, Zhu, NPSG and INLPG, performed poorly in all three change scenarios. Shi's overall performance is relatively balanced, with a KC of 77.35 in S1. Furthermore, Shi maintains a high KC value in S3, ranking second only to the proposed method. Unfortunately, Shi's KC value in S2 is significantly lower than CDIK, NR_ELM, ASEA, and the proposed method, with a KC value of 31.43%. The proposed method performs slightly worse than CDIK and Shi in S1, but achieves the best results in S2 and S3. Particularly in S2, the proposed algorithm achieves a KC evaluation value of 86.72%, exceeding CDIK by 10.77 percentage points.

[0091] Table 2 Evaluation values ​​of different change detection algorithms

[0092]

[0093] Example 4

[0094] A low-light wide-field-of-view image change detection device based on depth discrimination, the device includes: a processor and a memory, the memory storing program instructions, and the processor calling the program instructions stored in the memory to cause the device to execute the following method steps in Embodiment 1:

[0095] Training and validation datasets on object and noise variation features are created on bi-temporal image data used for training, and ON-DisNet is trained on them.

[0096] A global difference image is calculated on the test data, and the changed objects are retrieved according to a preset wide energy threshold; the retrieved changed objects are normalized and morphologically fused and filtered.

[0097] The trained ON-DisNet is used to determine whether the retrieved changing object is an object or noise, and local image pairs are obtained from the input bi-temporal test image based on the changes, orientation and scale information of the real object.

[0098] WeibDO is used to extract local difference features from local image pairs and create local difference images; LoG-Graph is applied to perform binary segmentation on the local difference images to generate local change maps, and these are then combined into a global change map.

[0099] Specifically, calculating the global difference image on the test data involves:

[0100]

[0101] in, and These are the mean values ​​of the input biphase sequence images T1 and T2, respectively.

[0102] The normalization and morphological fusion filtering of the retrieved changed objects are as follows:

[0103] Morphological filtering is performed on the normalized candidate objects using structuring elements E1, E2, and E3 to fuse and amplify the variation features:

[0104]

[0105] Where, ω l Represents the weighting coefficient, L represents the number of objects to be merged, and I represents the weighting coefficient. input Indicates the input image, I f This indicates the fusion result.

[0106] Specifically, WeibDO is used to extract local difference features from local image pairs as follows:

[0107]

[0108] Where k and λ are the shape parameter and scale parameter, respectively, and I represents the extreme pixel ratio. The calculation principle is as follows:

[0109]

[0110] Where I1 and I2 represent dual-phase input sequences, and min and max represent the minimum and maximum values ​​of the pixels to be retained for comparison, respectively.

[0111] The method also includes: using the LoG operator to construct the edge weights of the graph structure, the principle of which is:

[0112]

[0113] Where G represents the Gaussian function, and LoG represents the LoG operator. σ is the given standard deviation, set to 3.

[0114] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0115] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0116] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0117] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0118] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0119] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0120] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0121] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0122] References

[0123] [1]Singh A.Review articledigital change detection techniques usingremotely-sensed data[J].International journal of remote sensing,1989,10(6):989-1003.

[0124] [2]Bazi Y,Bruzzone L,Melgani F.An unsupervised approach based on thegeneralized Gaussianmodel to automatic change detection in multitemporal SARimages[J].IEEE Transactions onGeoscience and Remote Sensing,2005,43(4):874-887.

[0125] [3]Inglada J,Mercier G.A new statistical similarity measure forchange detection inmultitemporal SAR images and its extension to multiscalechange analysis[J].IEEE transactionson geoscience and remote sensing,2007,45(5):1432-1445.

[0126] [4]Gong M,Cao Y,Wu Q.A neighborhood-based ratio approach for changedetection in SARimages[J].IEEE Geoscience and Remote Sensing Letters,2011,9(2):307-311.

[0127] [5]Shi B,Jia Z,Yang J,et al.Unsupervised change detection in wide-field video images underlow illumination[J].IEEE Transactions on Circuits andSystems for Video Technology,2022,33(4):1564-1576.

[0128] [6]Celik T.Unsupervised change detection in satellite images usingprincipal component analysisand$k$-means clustering[J].IEEE geoscience andremote sensing letters,2009,6(4):772-776.[7]Zheng Y,Zhang X,Hou B,et al.Usingcombined difference image and$k$-means clusteringfor SAR image changedetection[J].IEEE Geoscience and Remote Sensing Letters,2013,11(3):691-695.

[0129] [8]Gao F,Dong J,Li B,et al.Change detection from synthetic apertureradar images based onneighborhood-based ratio and extreme learning machine[J].Journal of Applied Remote Sensing,2016,10(4):046019-046019.

[0130] [9]Gao F,Wang X,Gao Y,et al.Sea ice change detection in SAR imagesbased onconvolutional-wavelet neural networks[J].IEEE Geoscience and RemoteSensing Letters,2019,16(8):1240-1244.

[0131]

[10] Lv Z,Wang F J,Liu T,et al.Novel automatic approach for land coverchange detection byusing VHR remote sensing images[J].IEEE Geoscience andRemote Sensing Letters,2021,19:1-5.

[0132]

[11] Zhu Y,Jia Z,Yang J,et al.Change detection in multitemporalmonitoring images under lowillumination[J].IEEE Access,2020,8:126700-126712.

[0133]

[12] Sun Y,Lei L,Li X,et al.Nonlocal patch similarity basedheterogeneous remote sensingchange detection[J].Pattern Recognition,2021,109:107598.

[0134]

[13] Sun Y,Lei L,Li X,et al.Structure consistency-based graph forunsupervised changedetection with homogeneous and heterogeneous remotesensing images[J].IEEE transactions ongeoscience and remote sensing,2021,60:1-21.

[0135]

[14] Boykov Y,Kolmogorov V.An experimental comparison of min-cut / max-flow algorithms for energy minimization in vision[J].IEEE transactions onpattern analysis and machine intelligence,2004,26(9):1124-1137.

[0136] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0137] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0138] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-illumination wide-field image change detection method based on deep discriminant, characterized in that, The method includes: Training and validation datasets on object and noise variation features are created on bi-temporal image data used for training, and ON-DisNet is trained on the dataset. ON-DisNet is designed based on the lightweight residual network ResNet18. A global difference image is calculated on the test data, and the changed objects are retrieved according to a preset wide energy threshold; the retrieved changed objects are normalized and morphologically fused and filtered. The trained ON-DisNet is used to determine whether the retrieved changing object is an object or noise, and local image pairs are obtained from the input bi-temporal test image based on the changes, orientation and scale information of the real object. WeibDO is used to extract local difference features from local image pairs and create local difference images; LoG-Graph is applied to perform binary segmentation on the local difference images to generate local change maps, and these are then combined into a global change map. The normalization and morphological fusion filtering of the retrieved changed objects are as follows: Use structural elements , ,and Morphological filtering is applied to the normalized candidate objects to fuse and amplify the variation features. ； in, Indicates the weighting coefficient. Indicates the number of objects to be merged. Indicates the input image. Indicates the fusion result; The method of using WeibDO to extract local difference features from local image pairs is as follows: ； in, and These are shape parameters and scale parameters. Represents the extreme pixel ratio; calculation principle: ； in, and Indicates a two-phase input sequence. and These represent retaining the minimum and maximum values ​​among the compared pixels, respectively.

2. The low-light wide-field image change detection method based on depth discrimination according to claim 1, characterized in that, The calculation of the global difference image on the test data specifically involves: ； in, and These are the input biphase sequence images. and The mean.

3. The low-light wide-field image change detection method based on depth discrimination according to claim 1, characterized in that, The method further includes: using the LoG operator to construct edge weights for the graph structure, the principle of which is: ； wherein denotes a Gaussian function, denotes a LoG operator, is a given standard deviation.

4. A low-illumination wide-field image change detection device based on deep discriminant, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-3.

5. A computer readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-3.

Images