A remote sensing image apparent quality automatic detection method and system

Abstract

The application discloses a remote sensing image apparent quality automatic detection method and system, belongs to the field of remote sensing image processing, and solves the technical problems that the prior art is only concentrated on the detection of a certain apparent quality type, cannot quickly identify and regionally position various apparent quality abnormalities in large-scale remote sensing images, and has low real-time performance and precision in apparent quality detection. A data set is constructed, and diversified data enhancement is performed on the data set; after the diversified data enhanced data set is processed, the diversified data enhanced data set is used to train a target detection network; diversified data enhancement is performed on high-resolution remote sensing images, the trained target detection network is used to perform target detection on the diversified data enhanced high-resolution remote sensing images, and abnormal statistical results are obtained. The application is used for quickly identifying and regionally positioning various apparent quality abnormalities in large-scale remote sensing images.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing, in particular to a remote sensing image apparent quality automatic detection method and system. BACKGROUND

[0002] During the on-orbit operation of the optical remote sensing satellite based on the integration of space and ground, the remote sensing image may have different degrees of apparent quality problems due to the influence of factors such as changes in the space environment and fluctuations in sensor performance, especially in the aspects of large-scale, low-frequency gradual color difference and strip noise. Therefore, in order to ensure the accuracy of the evaluation of the apparent quality of the remote sensing image, it is necessary to continuously carry out apparent quality automatic detection on the image of the optical remote sensing satellite based on the integration of space and ground.

[0003] In recent years, deep learning technology has made significant progress in the field of remote sensing image analysis, especially in tasks such as target detection, image segmentation and anomaly detection, and has shown superior feature extraction and representation capabilities. Models based on convolutional neural networks can automatically learn multi-level image features, effectively capturing complex texture and structure information, so they are widely used in change detection, target recognition and other applications in remote sensing images.

[0004] However, the current method for detecting the apparent quality of remote sensing images is not mature. Existing researches mainly focus on the detection of a certain type of apparent quality, and lack a systematic and unified detection model that can adapt to various apparent quality anomalies. At the same time, how to balance the real-time and accuracy of detection to meet the demand for rapid processing of large-scale remote sensing data is also a technical problem that needs to be solved.

[0005] In summary, the existing technology has the technical problems that it only focuses on the detection of a certain type of apparent quality, cannot quickly identify and locate multiple apparent quality anomalies in large-scale remote sensing images, and has low real-time and accuracy in apparent quality detection. SUMMARY

[0006] The present application solves the technical problems of the prior art that only focuses on the detection of a certain type of apparent quality, cannot quickly identify and locate multiple apparent quality anomalies in large-scale remote sensing images, and has low real-time and accuracy in apparent quality detection.

[0007] The remote sensing image apparent quality automatic detection method according to the present application comprises the following steps:

[0008] Step 1, constructing a data set and performing diversified data enhancement on the data set;

[0009] Step 2, after processing the diversified data enhanced data set, using it to train the target detection network;

[0010] Step 3, the diversified data enhancement is carried out on the high-resolution remote sensing image, the trained target detection network is used for target detection on the diversified data enhanced high-resolution remote sensing image, and an abnormal statistical result is obtained.

[0011] Further, in an embodiment of the present application, the data set in step 1 is constructed, specifically:

[0012] The training samples are screened, and the multiple apparent quality anomaly categories existing in the screened training samples are labeled respectively, and the labeled training samples are used as the data set.

[0013] Further, in an embodiment of the present application, the diversified data enhancement in step 1 is specifically:

[0014] The data set is locally stretched based on a regional dynamic weight, the processed data set is resampled, and the diversified data enhanced data set is obtained.

[0015] Further, in an embodiment of the present application, the local stretching processing of the data set based on the regional dynamic weight is specifically:

[0016] The training samples in the data set are divided into multiple local blocks, the local gray range corresponding to each local block is obtained, each local block is linearly stretched respectively, the multiple linearly stretched local blocks are fused by weight, and the processed data set is obtained.

[0017] Further, in an embodiment of the present application, the training of the target detection network in step 2 is specifically:

[0018] Multiple apparent quality anomaly categories are set for the target detection network, the processed data set is used for cross-validation of the target detection network, and the trained target detection network is obtained.

[0019] Further, in an embodiment of the present application, the target detection on the diversified data enhanced high-resolution remote sensing image by using the trained target detection network in step 3 is specifically:

[0020] The high-resolution remote sensing image is uniformly divided into multiple sub-windows, and the trained target detection network is input into the trained target detection network after removing the overlapping prediction boxes for target detection.

[0021] Further, in an embodiment of the present application, the overlapping prediction boxes are removed, specifically:

[0022] The confidence of multiple overlapping prediction boxes in multiple adjacent sub-windows is calculated respectively, the overlapping prediction box with the highest confidence in each adjacent sub-window is retained, and the other overlapping prediction boxes in the adjacent sub-window are removed.

[0023] Further, in one embodiment of the present application, the multiple apparent quality anomaly categories include wide stripes, truncation, gradual change, inter-slice stripes and color difference caused by poor uniformity.

[0024] Further, in one embodiment of the present application, the target detection network further includes a transverse attention mechanism module, which is specifically:

[0025] The horizontal direction of the input image is globally averaged pooled, the globally averaged pooled result is channel compressed by a first convolutional layer, and after nonlinear activation, the channel is restored by a second convolutional layer, and the attention weight is obtained by processing through an activation function, the input image is weighted by using the attention weight, the horizontal position encoding of the input image is obtained, the dimension of the horizontal position encoding of the input image is expanded, the input image after feature weighting and the horizontal position encoding after dimension expansion are combined, and an enhanced image is output.

[0026] The remote sensing image apparent quality automatic detection system provided by the present application, the system is constructed according to the above method, including the following modules:

[0027] The data augmentation module constructs a data set and performs diversified data augmentation on the data set;

[0028] The model training module processes the diversified data augmented data set, and uses it to train the target detection network;

[0029] The anomaly detection module performs diversified data augmentation on the high-resolution remote sensing image, and uses the trained target detection network to perform target detection on the diversified data augmented high-resolution remote sensing image to obtain an anomaly statistical result.

[0030] The present application solves the technical problems in the prior art that the detection is only focused on a certain apparent quality type, cannot quickly identify and regionally locate multiple apparent quality anomalies in large-scale remote sensing images, and the real-time performance and accuracy of apparent quality detection are low. Specific beneficial effects include:

[0031] The application provides a remote sensing image apparent quality automatic detection method, which realizes high-precision automatic detection by constructing a multi-class radiation anomaly label system and using an advanced YOLOv12 target detection network. Meanwhile, a horizontal attention mechanism module is added to the YOLOv12 target detection network, which significantly improves the learning ability of color difference characteristics and further optimizes the detection precision. Combined with a sliding window block reasoning strategy, the target detection network can realize fast batch processing of remote sensing large-format images while maintaining high detection precision, meeting the real-time requirements of actual on-orbit satellite image quality control. Through various data enhancement technologies such as local stretching and resampling based on regional dynamic weights, the target detection network has good adaptability and robustness to remote sensing images of different sensors, various apparent quality anomaly forms and complex environments, and is suitable for multiple types of remote sensing platform images, and can realize fast identification and regional positioning of various apparent quality anomalies in large-scale remote sensing images.

[0032] The application combines regional recommendation mechanism and low-frequency color difference modeling to provide effective support for automatic production and quality inspection of remote sensing images, without human intervention, and has high automation, strong reliability, high timeliness and low cost. BRIEF DESCRIPTION OF DRAWINGS

[0033] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in conjunction with the drawings, in which:

[0034] Figure 1 is a flowchart of the remote sensing image apparent quality automatic detection method based on the integration of space and earth according to the first embodiment;

[0035] Figure 2 is a schematic diagram of the target detection network according to the fourth embodiment;

[0036] Figure 3 is a schematic diagram of the horizontal attention mechanism module according to the fourth embodiment;

[0037] Figure 4 is a schematic diagram of the gradual color difference detection position and confidence in the port scene according to the application;

[0038] Figure 5 is a schematic diagram of the wide stripe detection position and confidence in the all-cloud scene according to the application;

[0039] Figure 6 is a schematic diagram of the tap problem detection position and confidence in the sea surface scene according to the application;

[0040] Figure 7 is a schematic diagram of the gradual color difference detection position and confidence in the mountainous area scene according to the application. DETAILED DESCRIPTION

[0041] The various embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings. The embodiments described by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0042] Embodiment one. A remote sensing image apparent quality automatic detection method, comprising the following steps:

[0043] Step 1, constructing a data set and diversifying data enhancement on the data set;

[0044] Step 2, after processing the diversified data enhanced data set, training a target detection network using the same;

[0045] Step 3, diversifying data enhancement on high-resolution remote sensing images, using the trained target detection network to detect targets on the diversified data enhanced high-resolution remote sensing images, and obtaining abnormal statistical results.

[0046] The prior art has the problem that research is mainly focused on detection of a certain apparent quality type, resulting in lack of a systematic, unified detection model suitable for diversified apparent quality abnormalities, and inability to balance real-time and precision of detection, meeting the demand for rapid processing of large-scale remote sensing data.

[0047] To solve the above technical problems, the present embodiment provides a remote sensing image apparent quality automatic detection method, as shown in Figure 1 The data set in step 1 is mainly from real remote sensing data taken by Jilin No. 1 satellite constellation, to meet the high-precision training needs of YOLO (a target detection algorithm based on deep learning) series target detection model in apparent quality problem automatic identification, and the target detection network in step 3 uses YOLOv12 as the target detection network, normalizes the input data to ensure that the input data meets the input specification of the YOLO detection network, and improves the saliency of apparent quality abnormal features.

[0048] Therefore, the present embodiment constructs a data set and diversifies data enhancement on the data set, processes the diversified data enhanced data set, trains a target detection network using the same, uses the trained target detection network to detect targets on the diversified data enhanced high-resolution remote sensing images, and obtains abnormal statistical results, solving the technical problems in the prior art that only focus on detection of a certain apparent quality type, cannot quickly identify and regionally locate multiple apparent quality abnormalities in large-scale remote sensing images, and the real-time and precision of apparent quality detection are low.

[0049] Embodiment two. The embodiment is a further limitation of the remote sensing image apparent quality automatic detection method described in embodiment one. The step 1 of constructing a data set is specifically:

[0050] The training samples are screened, and the multiple apparent quality abnormal categories existing in the screened training samples are labeled respectively. The labeled training samples are used as the data set.

[0051] In this embodiment, the data set mainly includes multiple common apparent quality abnormal types in remote sensing images, such as wide stripes, truncation, gradual change, inter-slice stripes, and color difference caused by poor uniformity. The training samples are manually screened and labeled by experts to ensure the representativeness of the data and the accuracy of the problem labeling. According to the input requirements of the YOLO detection network, all apparent quality abnormal regions in each training sample are labeled in a standard bounding box format during manual labeling. The labeling adopts the YOLO detection network format to ensure that the labeling box and the apparent quality abnormal category label standard are unified, facilitating efficient calling during the training and reasoning processes.

[0052] Embodiment three. The embodiment is a further limitation of the remote sensing image apparent quality automatic detection method described in embodiment one. The diversified data enhancement in step 2 is specifically:

[0053] The data set is locally stretched based on regional dynamic weights, and the processed data set is resampled to obtain a diversified data enhanced data set.

[0054] In this embodiment, the data set is locally stretched based on regional dynamic weights, which is specifically:

[0055] The training samples in the data set are divided into multiple local blocks, the local gray range corresponding to each local block is obtained, each local block is linearly stretched, the multiple linearly stretched local blocks are weight fused, and the processed data set is obtained.

[0056] Since remote sensing images are often stored in high-bit depth format to fully preserve the dynamic range of the radiation response of ground objects, in order to facilitate manual labeling of radiation abnormalities after detection and to ensure that the network model input is suitable, this embodiment performs local stretching based on regional dynamic weights on all original remote sensing images.

[0057] First, the local stretching function is calculated in blocks. Suppose the original remote sensing image is divided into local blocks, each local block is , where , , is the number of horizontal blocks after the original remote sensing image is divided, The number of column blocks after dividing the original remote sensing image. Calculate the local grayscale range of each local block:

[0058] ;

[0059] ;

[0060] in, For local blocks Minimum value within, For local blocks The maximum value within, for Image values. Then, intra-block linear stretching is performed, and the stretched values ​​for each local block are:

[0061] ;

[0062] in, for The values ​​are obtained after stretching the internal image. Then, weighted fusion is performed on multiple local blocks after linear stretching, i.e., weighted fusion is performed on the entire image to avoid block artifacts. Each pixel... The final value is the weighted average of its local block, with Gaussian weights. The weighted image is inversely proportional to the distance from the pixel to the center of the local block. The resulting remote sensing image after weighted fusion is:

[0063] ;

[0064] The Gaussian weights are defined as follows:

[0065] ;

[0066] in,( , ) is a local block The center, is the Gaussian kernel standard deviation.

[0067] This implementation effectively compresses and maps the image grayscale range to an 8-bit image, which can highlight abnormal features, improve annotation efficiency, and maintain the relative contrast of the original radiometric characteristics.

[0068] YOLO detection networks typically require input images of a fixed size to standardize the network structure and accelerate inference efficiency. Since the original dimensions of remote sensing images are often non-standard resolutions, this implementation performs a resampling operation after image stretching.

[0069] Implementation Method 4. This implementation method further defines the automated remote sensing image appearance quality detection method described in Implementation Method 1. Specifically, training the target detection network in step 3 involves:

[0070] The target detection network is set with multiple apparent quality anomaly categories, and cross-validation is performed on the target detection network using the processed data set to obtain a trained target detection network.

[0071] In this embodiment, the target detection network further comprises a transverse attention mechanism module, which specifically comprises:

[0072] The horizontal direction of the input image is globally averaged pooled, the globally averaged pooled result is channel compressed by a first convolutional layer, and after nonlinear activation, the channel is restored by a second convolutional layer, and the attention weight is obtained by processing through an activation function. The input image is weighted using the attention weight, the horizontal position encoding of the input image is obtained, the dimension of the horizontal position encoding of the input image is expanded, the input image after feature weighting and the horizontal position encoding after dimension expansion are combined, and an enhanced image is output.

[0073] In this embodiment, based on a diversified and completely labeled radiation anomaly data set, an end-to-end target detection network is trained. Cross-validation and index evaluation are used to evaluate the performance of the model to ensure its detection accuracy and stability on actual remote sensing images. To ensure the effectiveness and generalization ability of the training, the processed data set is divided into a training set and a validation set in a ratio of 9:1. The distribution of the number of samples of each class is kept balanced during the division process to avoid training bias. The division method uses stratified sampling to ensure that each class of apparent quality anomaly has representative samples in both subsets.

[0074] This embodiment uses YOLOv12 as the target detection network. The YOLOv12 target detection network integrates advanced Transformer (self-attention mechanism-based model) attention mechanism and improved backbone network, enhancing the feature extraction capability. The network includes multi-level feature fusion modules, improving the response capability to different size anomaly targets, such as Figure 2 , wherein the SiLU function is an activation function, P3, P4 and P5 are three detection layers, the C3K2 module is inherited from YOLOv11, using CSP architecture and depth separable convolution for auxiliary feature extraction, the A2C2f module is a regional attention and two CSP bottleneck layers for feature extraction, further processing and refining features.

[0075] This embodiment is aimed at the transverse gradual color difference detection task and has made targeted improvements based on the YOLOv12 target detection network, adding a transverse attention mechanism module, as shown in Figure 2 and Figure 3 The core of the transverse attention mechanism module is to enhance the sensitivity of the model to the horizontal direction ( axis) feature change, while suppressing the vertical direction To eliminate the irrelevant interference of the vertical direction (height) and the channel dimension, the horizontal spatial information is compressed by global average pooling in the horizontal direction (width) of the feature map:

[0076] 1. Horizontal spatial weight distribution: Global pooling is performed on the horizontal direction of the input feature map to generate attention weights for each image coordinate, highlighting areas with dramatic changes in the horizontal direction. Specifically:

[0077] To preserve the vertical direction (height ) and channel dimension, the horizontal spatial information is compressed by global average pooling in the horizontal direction (width ) of the feature map:

[0078] ;

[0079] where is the pooled image, is the 2D average pooling, is the original feature map, is the kernel size, set to (pooling only in the width direction).

[0080] Attention Weight Calculation (Attention Weight Generation) is performed by two layers of convolution (including nonlinear activation) on the global average pooling result to compress and restore the channel, generating attention weights matching the input feature map channel number :

[0081] ;

[0082] where is the activation function, is the convolution layer, is the ReLu (Rectified Linear Unit) nonlinear activation function.

[0083] The attention weights are multiplied by the original feature map element by element to perform feature weighting and enhance the horizontal direction key features:

[0084] ;

[0085] where is the input feature map after feature weighting, is the Hadamard product.

[0086] 2. Horizontal position encoding: The horizontal coordinate information is converted into features and fused with the original features to allow the target detection network to perceive the association between "horizontal position" and "color features". Specifically:

[0087] First, the horizontal position encoding generation is performed, and the horizontal coordinate is encoded by using the sine and cosine encoding, and the encoding formula is:

[0088]

[0089]

[0090] wherein, is the value of the horizontal position encoding on the horizontal coordinate , the channel , is the horizontal coordinate, , and the value is 0 to , is the channel index, , and the value is 0 to .

[0091] Then, the encoding is fused with the features. The horizontal position encoding is extended to the same dimension as the (batch and height direction replication), and is added to :

[0092] ;

[0093] wherein, is the enhanced image output by the transverse attention module, is the extended horizontal position encoding, and the batch dimension is replicated times and the height dimension is replicated times.

[0094] In summary, the complete output of the transverse attention module is summarized as:

[0095] ;

[0096] In this embodiment, in view of the target characteristics of the apparent quality anomaly, the anchor box size is adjusted in combination with the detection head of the YOLOv12 target detection network which is suitable for multi-size targets, so as to optimize the detection performance and improve the confidence of quality problem classification and the accuracy of the position of the abnormal area. According to the typical apparent quality anomaly types in remote sensing images, a plurality of apparent quality anomaly categories are set, i.e., wide stripes, truncation, gradual change, inter-slice stripes, and color difference caused by poor uniformity.

[0097] In this embodiment, a comprehensive loss function is used, which includes a boundary box regression loss, a category loss and a confidence loss. Specifically, the boundary box regression loss ensures the accuracy of the position and size of the frame, and the formula is as follows:

[0098] ;

[0099] wherein, is the intersection over union of the predicted box ( ) and the ground truth box ( ), is the squared Euclidean distance between the centers of the predicted box and the ground truth box, is the length of the diagonal of the minimum enclosing rectangle that encloses both boxes, , are the squared width and height difference of the predicted box and the ground truth box, respectively, , are the width and height of the image, respectively.

[0100] The class loss is used to ensure the correct classification, and the specific formula is as follows:

[0101] ;

[0102] wherein, is the true class label, is the class probability predicted by the target detection network, is the dynamic class weight, is the dynamic focus parameter.

[0103] The confidence loss is used to ensure that the model can distinguish between the target and the background, and the formula is as follows:

[0104]

[0105] wherein, is the true confidence, is the confidence predicted by the model, is the IoU of the predicted box and the ground truth box as the weight.

[0106] Further, in the present embodiment, the Adam optimizer is used in the training, the learning rate strategy is dynamically adjusted to accelerate the convergence, and the early stopping mechanism is introduced to prevent overfitting.

[0107] Embodiment Five. The present embodiment is a further limitation of the remote sensing image apparent quality automatic detection method described in Embodiment One. In step 3, the trained target detection network is used to detect the high-resolution remote sensing image after data augmentation, and the specific steps are as follows:

[0108] The high-resolution remote sensing image is uniformly divided into multiple sub-windows, and after removing the overlapping predicted boxes, it is input into the trained target detection network for target detection.

[0109] ​​The overlapping prediction frame is removed according to the embodiment, and for calculating the confidence of the plurality of overlapping prediction frames in the plurality of adjacent sub-windows, the overlapping prediction frame with the highest confidence in each adjacent sub-window is reserved, and the other overlapping prediction frames in the adjacent sub-window are removed.

[0110] In order to adapt to high-resolution remote sensing images, and to realize fast batch processing of remote sensing large-format images while maintaining high detection accuracy, and to meet the real-time demand of actual on-orbit satellite image quality control, the sliding window block reasoning is adopted in the embodiment, the non-maximum suppression algorithm is combined to remove the repeated frame, the high-resolution remote sensing image is divided into a plurality of sub-windows with fixed size, and the trained target detection network is input block by block for target detection.

[0111] However, when the window slides, a certain overlapping area is reserved, in order to avoid missing detection due to the target being segmented at the block edge, but the overlapping area of adjacent sub-windows may detect the same target repeatedly, resulting in multiple overlapping prediction frames. In order to solve the above problems, the IoU (intersection over union) of the overlapping prediction frames is calculated by the NMS (non-maximum suppression) algorithm, the overlapping prediction frame with the highest confidence is reserved, the redundant overlapping prediction frames highly overlapped with it are deleted, and finally the unique target frame is obtained. The embodiment guarantees the detection efficiency and accuracy of the target detection network on large-scale images, and meets the demand of automatic detection of apparent quality.

[0112] In summary, the trained target detection network has good adaptability and robustness to remote sensing images of different sensors, various apparent quality abnormal patterns and complex environments, and is suitable for images of multiple types of remote sensing platforms, such as Figure 4 、 Figure 5 、 Figure 6 and Figure 7 As shown, the detection position and confidence diagram of the gradual color difference in the port scene, the wide stripe in the all-cloud scene, the tap problem in the sea surface scene and the gradual color difference in the mountain scene are shown, and the present application can realize fast identification and regional positioning of various apparent quality abnormalities in large-scale remote sensing images.

[0113] Embodiment six. The remote sensing image apparent quality automatic detection system according to the embodiment, the system is constructed according to the method of embodiment one, comprising the following modules:

[0114] The data enhancement module constructs a data set and performs diversified data enhancement on the data set;

[0115] The model training module processes the diversified data enhanced data set, and trains the target detection network using the data set;

[0116] Anomaly detection module performs diversified data enhancement on the high-resolution remote sensing image, and uses the trained target detection network to perform target detection on the diversified data enhanced high-resolution remote sensing image to obtain an anomaly statistical result.

[0117] The above describes in detail the remote sensing image apparent quality automatic detection method and system provided by the present application, and the principles and implementation manners of the present application are described by using specific examples. The above example is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for automatic detection of apparent quality of remote sensing images, characterized in that, The method comprises the following steps: Step 1, constructing a data set and performing diversified data enhancement on the data set; Step 2, processing the diversified data enhanced data set, and training a target detection network by using the data set; Step 3, performing diversified data enhancement on the high-resolution remote sensing image, performing target detection on the diversified data enhanced high-resolution remote sensing image by using the trained target detection network, and obtaining an abnormal statistical result; The diversified data enhancement in step 1 is specifically: Performing local stretching processing on the data set based on a regional dynamic weight, resampling the processed data set, and obtaining a diversified data enhanced data set; The local stretching processing on the data set based on the regional dynamic weight is specifically: Dividing the training samples in the data set into a plurality of local blocks, obtaining a local gray scale range corresponding to each local block, performing linear stretching on each local block respectively, performing weight fusion on the linearly stretched plurality of local blocks, and obtaining a processed data set; The target detection in step 3 by using the trained target detection network on the diversified data enhanced high-resolution remote sensing image is specifically: Uniformly dividing the high-resolution remote sensing image into a plurality of sub-windows, removing overlapping prediction boxes, and inputting the high-resolution remote sensing image into the trained target detection network for target detection; The removing of the overlapping prediction boxes is specifically: Calculating the confidence of a plurality of overlapping prediction boxes in a plurality of adjacent sub-windows respectively, retaining the overlapping prediction box with the highest confidence in each adjacent sub-window, and removing other overlapping prediction boxes in the adjacent sub-window; The target detection network further comprises a horizontal attention mechanism module, and the horizontal attention mechanism module is specifically: Performing global average pooling on the horizontal direction of an input image, performing channel compression on the global average pooling result by using a first convolutional layer, performing feature weighting on the input image by using the attention weight, obtaining the horizontal position encoding of the input image, expanding the dimension of the horizontal position encoding of the input image, combining the input image after the feature weighting and the horizontal position encoding after the dimension expansion, and outputting an enhanced image.

2. The method of claim 1, wherein the method further comprises: The constructing of the data set in step 1 is specifically: Screening the training samples, labeling a plurality of apparent quality anomaly categories existing in the screened training samples respectively, and taking the labeled training samples as a data set.

3. The method of claim 1, wherein the method further comprises: The training of the target detection network in step 2 is specifically: Setting a plurality of apparent quality anomaly categories for the target detection network, performing cross-validation on the target detection network by using the processed data set, and obtaining a trained target detection network.

4. The method according to claim 2 or 3, characterized in that, The plurality of apparent quality anomaly categories include wide stripes, truncation, gradual change, inter-slice stripes and color difference caused by poor uniformity.

5. A system for automatic detection of apparent quality of remote sensing imagery, said system being constructed in accordance with the method of claim 1, characterized by, The method comprises the following modules: A data enhancement module, which constructs a data set and performs diversified data enhancement on the data set; A model training module, which trains a target detection network by using the diversified data enhanced data set after processing the data set; A target detection network trained by using the diversified data enhanced data set after processing the data set; Anomaly detection module, the high-resolution remote sensing image is diversified data enhancement, the target detection network of training completion is used to the high-resolution remote sensing image of diversified data enhancement for target detection, and the abnormal statistical result is obtained. Anomaly detection module, the high-resolution remote sensing image is diversified data enhancement, the target detection network of training completion is used to the high-resolution remote sensing image of diversified data enhancement for target detection, and the abnormal statistical result is obtained.

Images