A satellite image target recognition method and system based on deep learning
Through deep learning methods, including data annotation, segmentation scaling and model training, combined with ResNet-152 network, the problem of inaccurate object recognition in satellite images is solved, the recognition accuracy and reliability are improved, and the effectiveness of large-scale satellite images is achieved Identify.
Patent Information
- Application Number
- CN202011231793.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-11-06
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-11-06
AI Technical Summary
Traditional satellite image object detection and classification algorithms have problems of inaccuracy and unreliability, especially because the size of objects and facilities in satellite images is large, resulting in the loss of details after the image is reduced, affecting the recognition accuracy.
Using a deep learning-based method, data annotation, segmentation scaling, data division amplification and model training are used, including transforming the annotation frame from a rectangle to a square to expand the annotation range, segmenting and scaling the data set, rotating and color change amplification, And iterative optimization was performed using the ResNet-152 network model, and the recognition effect was evaluated in combination with F1-Score.
The accuracy and reliability of satellite map target recognition are improved, and the recognition accuracy is improved by retaining context information and confusing data, and the problem that large satellite maps cannot be directly identified is solved.
Smart Images

Figure CN112232288B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image recognition processing, and specifically relates to a satellite image target recognition method and system based on deep learning. Background Art
[0002] Finding specific objects, facilities, and events in satellite imagery is an important problem, and satellite imagery needs to be effectively preprocessed for input into deep networks. To keep processing time reasonable, deep learning requires relatively small images of fixed size, however, this is not the case with satellite imagery, as objects and facilities can be much larger than those in regular photographs. When the image size is compressed, small details in the image disappear. These details may include important distinguishing features.
[0003] Problems or defects in existing technologies: Traditional object detection and classification algorithms are inaccurate and unreliable. Summary of the invention
[0004] In view of the technical problems of inaccuracy and unreliability in the above-mentioned traditional target detection and classification algorithms, the present invention provides a satellite image target recognition method and system based on deep learning with high accuracy, strong reliability and good recognition effect.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0006] A satellite image target recognition method based on deep learning includes the following steps:
[0007] S1. Data labeling: Label the objects to be identified in the satellite image. The labeling information includes the category of the object, the location of the object, and the easily confused objects. At the same time, the labeling box is changed from a rectangle to a square, the labeling range is expanded, and the context information around the target to be identified is included in the labeling range;
[0008] S2, data segmentation and scaling: segment the marked objects in the satellite image according to their location information. The segmentation content includes the marked information, part of the data at the edge of the marked object, and the easily confused objects. Together with the image of the object to be identified, the data set is formed and the data size of the data set is scaled according to the scale of the convolutional neural network input image.
[0009] S3, data division and amplification: divide the data set into training set, validation set and test set, and rotate, change the color and increase the data quantity of the training set;
[0010] S4. Model training: Input the assembled training set and its corresponding labels into the network, and train the model through iterative optimization of the network. When the model training loss no longer decreases, use the validation set data to continue training the model. If the model loss no longer decreases, save the model. If the loss continues to decrease, use the training set to continue iteratively training the model until the loss value stabilizes. Save the model, input the test set data into the trained model for recognition, and evaluate the recognition results.
[0011] The calculation method of data annotation in S1 is as follows: mark the n types of objects to be identified in the satellite image, and mark them as 1, 2...n, and also mark the objects similar to the objects to be identified. The similar objects are uniformly marked as 0. When marking, the location information of the target object is also marked. The object is marked by a rectangular box. Therefore, the annotation information of each object contains 5 information (l 0 ,l 1 ,l 2 ,l 3 ,l 4 ), l 0 is the category information, (l 1 ,l 2 ,l 3 ,l 4 ) is the location information, and the four points correspond to the upper left, lower left, lower right, and upper right respectively. The maximum length of the object position marking box is L=max(l 2 -l 1 ,l 3 -l 2 ) as the edge, convert the annotation box into a square, and then expand it by 10% to capture more edge and context information, provide more recognition features to the model, and then update its position annotation information. The new annotation information is (l 0 ,k 1 ,k 2 ,k 3 ,k 4 ).
[0012] The method for data segmentation and scaling in S2 is: the labeled data in S1 are divided according to their position information (k 1 ,k 2 ,k 3 ,k 4 ) to split and save, and the split data is sorted according to its corresponding category label l 0 Save and scale the data so that it can be input into the model, and adjust it to 224*224 in proportion.
[0013] The method for data division and amplification in S3 is: divide the data set into a training set, a validation set and a test set in a ratio of 7:2:1, the training set is used for model training, the validation set is used to detect whether the model loss continues to decrease, and the test set is used to test the model effect. The training set data is rotated 45 degrees, 90 degrees, and 135 degrees respectively, and the data contrast and brightness are adjusted to amplify the data, and the transformed data is mixed with the original training set to construct a new data set.
[0014] The model training method in S4 is as follows: the convolution kernel size of the first convolution is 7*7, and the subsequent convolutions are all 3*3. Except for the first layer, every two layers of the subsequent convolution layers constitute a residual block. The last layer is a fully connected layer for outputting the prediction results. The prediction results are used to obtain the prediction probability for each category through the sigmoid function. The network input size is 224*224. The processed training set and its corresponding label are input into the network to train the model, and the validation set data is used to train the model to see whether the model loss function continues to decrease. If it continues to decrease, it indicates that the model has not reached the best. The model is continued to be trained. If it no longer decreases, the model is saved. The model is used to test the test set data and compared with its label result. The result output by the model is the probability that the data is each data category. If the probability of each data category is greater than 0.5, it means that this data is of this category. If the probability of each data category is less than 0.5, it indicates that this data does not contain the object to be identified. The recognition effect is evaluated using F1-Score.
[0015] The evaluation method of the F1-Score is:
[0016]
[0017]
[0018]
[0019] The F 1 For recognition effect, TP is the correct prediction answer, FP is the wrong prediction of other classes as this class, FN is the prediction of this class label as other class labels, precision is the proportion of positive samples in the positive examples determined by the classifier, and recall is the proportion of positive samples predicted as positive examples in the total positive examples;
[0020] If the model recognition effect F 1 If the specified F1-Score value is reached, the model training is completed. 1 If the specified F1-Score value is not reached, adjust the model parameters and continue model training. The F1-Score value is the harmonic mean of the precision and recall, with a maximum of 1 and a minimum of 0.
[0021] A satellite image target recognition system based on deep learning includes a data annotation module, a data segmentation and scaling module, a data division and amplification module, and a model training module. The data annotation module is connected to the data segmentation and scaling module via communication, the data segmentation and scaling module is connected to the data segmentation and amplification module via communication, and the data segmentation and amplification module is connected to the model training module via communication.
[0022] The data segmentation and scaling module includes a data segmentation module and a data scaling module. The data segmentation module is connected to the data scaling module, the data segmentation module is connected to the data annotation module, and the data scaling module is connected to the data division and amplification module.
[0023] The data partitioning and amplification module includes a data partitioning module and a data amplification module. The data partitioning module is connected to the data amplification module. The data partitioning module is connected to the data scaling module of the data segmentation and scaling module. The data amplification module is connected to the model training module.
[0024] The model training module includes a training module, a recognition module, and an evaluation module. The training module is connected to the data amplification module of the data partitioning and amplification module. The training module is connected to the recognition module, and the recognition module is connected to the evaluation module. The model in the training module adopts the ResNet-152 network model, and the recognition module adopts F1-Score.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] The present invention solves the problem that large satellite images cannot be directly recognized by training the units to be recognized in large satellite images separately. At the same time, when segmenting the images of objects to be recognized, the context information is retained to a certain extent, which provides more features for network recognition. At the same time, easily confused data is added to help the network distinguish similar objects, thereby improving the recognition accuracy. The deep network model ResNet-152 is used to ensure the recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a work flow chart of the present invention;
[0028] Figure 2 It is a structural schematic diagram of the present invention;
[0029] Figure 3 This is a flow chart of model training of the present invention.
[0030] Among them: 1 is a data labeling module, 2 is a data segmentation and scaling module, 3 is a data division and amplification module, 4 is a model training module, 201 is a data segmentation module, 202 is a data scaling module, 301 is a data segmentation module, 302 is a data amplification module, 401 is a training module, 402 is a recognition module, and 403 is an evaluation module. DETAILED DESCRIPTION
[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] A satellite image target recognition method based on deep learning, such as Figure 1 As shown, the following steps are included:
[0033] Step 1: Data labeling: Label the objects to be identified in the satellite image. The labeling information includes the category of the object, the location of the object, and the easily confused objects. At the same time, the labeling box is changed from a rectangle to a square, the labeling range is expanded, and the contextual information around the target to be identified is included in the labeling range;
[0034] Step 2: Data segmentation and scaling: Segment the marked objects in the satellite image according to their location information. The segmentation content includes the marked information, some data on the edges of the marked objects, and easily confused objects. Together with the images of the objects to be identified, the dataset is formed and the data size of the dataset is scaled according to the scale of the convolutional neural network input image.
[0035] Step 3: Data division and amplification: Divide the data set into a training set, a validation set, and a test set, and rotate, change the color, and increase the amount of data in the training set;
[0036] Step 4: Model training: Input the assembled training set and its corresponding labels into the network, and train the model through iterative optimization of the network. When the model training loss no longer decreases, use the validation set data to continue training the model. If the model loss no longer decreases, save the model. If the loss continues to decrease, use the training set to continue iteratively training the model until the loss value stabilizes. Save the model, input the test set data into the trained model for recognition, and evaluate the recognition results.
[0037] Furthermore, the calculation method of data annotation in step 1 is: mark the n types of objects to be identified in the satellite image, and mark them as 1, 2...n, and also mark the objects similar to the objects to be identified. The similar objects are uniformly marked as 0. When marking, the location information of the target object is also marked. The object is marked by a rectangular box. Therefore, the annotation information of each object contains 5 information (l 0 ,l 1 ,l 2 ,l 3 ,l 4 ), l 0 is the category information, (l 1 ,l 2 ,l 3 ,l 4 ) is the location information, and the four points correspond to the upper left, lower left, lower right, and upper right respectively. The maximum length of the object position marking box is L=max(l 2 -l 1 ,l 3 -l 2 ) as the edge, convert the annotation box into a square, and then expand it by 10% to capture more edge and context information, provide more recognition features to the model, and then update its position annotation information. The new annotation information is (l 0 ,k 1 ,k 2 ,k 3 ,k 4 ).
[0038] Furthermore, the method of data segmentation and scaling in step 2 is: the data marked in step 1 are divided according to their position information (k 1 ,k 2 ,k 3 ,k 4 ) to split and save, and the split data is sorted according to its corresponding category label l 0 Save and scale the data so that it can be input into the model, and adjust it to 224*224 in proportion.
[0039] Furthermore, the method of data division and amplification in step 3 is: divide the data set into training set, validation set and test set in a ratio of 7:2:1, the training set is used for model training, the validation set is used to detect whether the model loss continues to decrease, and the test set is used to test the model effect. The training set data is rotated 45 degrees, 90 degrees, and 135 degrees respectively, and the data contrast and brightness are adjusted to amplify the data, and the transformed data is mixed with the original training set to construct a new data set.
[0040] Further, if Figure 3As shown in the figure, the model training method in step 4 is as follows: the convolution kernel size of the first convolution is 7*7, and the subsequent convolutions are all 3*3. Except for the first layer, every two convolution layers form a residual block. The last layer is a fully connected layer, which is used to output the prediction results. The prediction results are obtained by the sigmoid function to obtain the prediction probability for each category. The network input size is 224*224. The processed training set and its corresponding label are input into the network to train the model, and the validation set data is used to train the model to see whether the model loss function continues to decrease. If it continues to decrease, it indicates that the model has not reached the best. The model will continue to be trained. If it no longer decreases, the model will be saved. The model will be used to test the test set data and compared with its label results. The result of the model output is the probability that the data is each data category. If the probability of each data category is greater than 0.5, it means that this data is of this category. If the probability of each data category is less than 0.5, it indicates that this data does not contain the object to be recognized. The recognition effect is evaluated using F1-Score.
[0041] Furthermore, the evaluation method of F1-Score is:
[0042]
[0043]
[0044]
[0045] F 1 For recognition effect, TP is the correct prediction answer, FP is the wrong prediction of other classes as this class, FN is the prediction of this class label as other class labels, precision is the proportion of positive samples in the positive examples determined by the classifier, and recall is the proportion of positive samples predicted as positive examples in the total positive examples;
[0046] If the model recognition effect F 1 If the specified F1-Score value is reached, the model training is completed. 1 If the specified F1-Score value is not reached, adjust the model parameters and continue model training. The F1-Score value is the harmonic mean of precision and recall, with a maximum of 1 and a minimum of 0.
[0047] A satellite image target recognition system based on deep learning, such as Figure 2As shown, it includes a data annotation module 1, a data segmentation and scaling module 2, a data division and amplification module 3, and a model training module 4. The data annotation module 1 is connected to the data segmentation and scaling module 2 through communication. The data segmentation and scaling module 2 is connected to the data segmentation and amplification module 3 through communication. The data segmentation and amplification module 3 is connected to the model training module 4 through communication. The data annotation module 1 annotates the objects to be identified in the satellite image, transforms the annotation box from a rectangle to a square, and expands the annotation range, and includes the context information around the target to be identified in the annotation range; the data segmentation and scaling module 2 segments the objects annotated in the satellite image according to their location information, and scales the data size of the data set according to the scale of the convolutional neural network input image; the data segmentation and amplification module 3 divides the data set into a training set, a validation set, and a test set in proportion, rotates the data of the training set, changes the color, and expands the number of data; the model training module 4 inputs the assembled training set and its corresponding label into the network, trains the model through iterative optimization of the network, and uses the validation set data to train the model, inputs the test set data into the trained model for recognition, and evaluates the recognition result.
[0048] Furthermore, the data segmentation and scaling module 2 includes a data segmentation module 201 and a data scaling module 202 . The data segmentation module 201 is connected to the data scaling module 202 . The data segmentation module 201 is connected to the data labeling module 1 . The data scaling module 202 is connected to the data division and amplification module 3 .
[0049] Furthermore, the data partitioning and amplification module 3 includes a data partitioning module 301 and a data amplification module 302. The data partitioning module 301 is connected to the data amplification module 302. The data partitioning module 301 is connected to the data scaling module 202 of the data segmentation and scaling module 2. The data amplification module 302 is connected to the model training module 4.
[0050] Furthermore, the model training module 4 includes a training module 401, a recognition module 402, and an evaluation module 403. The training module 401 is connected to the data amplification module 302 of the data partitioning and amplification module 3. The training module 401 is connected to the recognition module 402. The recognition module 402 is connected to the evaluation module 403. The model in the training module 401 adopts the ResNet-152 network model, and the recognition module 402 adopts F1-Score.
[0051] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc., which contain computer-usable program codes.
[0052] Only the preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention, and various changes should be included in the protection scope of the present invention.
Claims
1. A satellite image target recognition method based on deep learning, characterized in that: The following steps are involved: S1. Data annotation; The objects to be identified in the satellite image are marked, and the marking information includes the category, location and easily confused objects. At the same time, the marking box is changed from a rectangle to a square, the marking range is expanded, and the context information around the target to be identified is included in the marking range; the calculation method of the data marking in S1 is: mark the n types of objects to be identified in the satellite image, marked as 1, 2...n, and mark the objects similar to the objects to be identified, and the similar objects are uniformly marked as 0. When marking, the location information of the target object is also marked. The object is marked by a rectangular box, so each object The annotation information contains 5 pieces of information (l0, l1, l2, l3, l4), l0 is the category information, (l1, l2, l3, l4) is the position information, and the four points correspond to the upper left, lower left, lower right, and upper right respectively. The maximum length of the object position annotation box L = max (l2-l1, l3-l2) is used as the edge, and the annotation box is converted into a square, and then it is enlarged by 10% to capture more edge and context information, providing more recognition features to the model, and then its position annotation information is updated, and the new annotation information is (l0, k1, k2, k3, k4); S2, data segmentation and scaling: segment the marked objects in the satellite image according to their location information. The segmentation content includes the marked information, part of the data at the edge of the marked object, and the easily confused objects. Together with the image of the object to be identified, the data set is formed and the data size of the data set is scaled according to the scale of the convolutional neural network input image. S3, data partitioning and expansion; The data set is divided into training set, validation set and test set, and the data in the training set is rotated, color changed, and the amount of data is increased; S4, model training; input the assembled training set and its corresponding labels into the network, train the model through iterative optimization of the network, and when the model training loss no longer decreases, continue to train the model using the validation set data. If the model loss no longer decreases, save the model. If the loss continues to decrease, continue to iteratively train the model using the training set until the loss value stabilizes, save the model, input the test set data into the trained model for recognition, and evaluate the recognition results; The model training method in S4 is as follows: the convolution kernel size of the first convolution is 7*7, and the convolutions thereafter are all 3*3. Except for the first layer, every two convolution layers thereafter constitute a residual block. The last layer is a fully connected layer for outputting the prediction results. The prediction results are obtained by the sigmoid function to obtain the prediction probability for each category. The network input size is 224*224. The processed training set and its corresponding label are input into the network to train the model, and the validation set data is used to train the model to see whether the model loss function continues to decrease. If it continues to decrease, it indicates that the model has not reached the best. The model is continued to be trained. If it no longer decreases, the model is saved. The model is used to test the test set data and compared with its label result. The result output by the model is the probability that the data belongs to each data category. If the probability of each data category is greater than 0.5, it indicates that this data is of this category. If the probability of each data category is less than 0.5, it indicates that this data does not contain the object to be identified. The recognition effect is evaluated using F1-Score. The evaluation method of the F1-Score is: F1 is the recognition effect, TP is the correct prediction answer, FP is the wrong prediction of other classes as this class, FN is the prediction of this class label as other class labels, precision is the proportion of positive samples in the positive examples determined by the classifier, and recall is the proportion of positive samples predicted as positive examples in the total positive examples; If the model recognition effect F1 reaches the specified F1-Score value, the model training is completed. If the model recognition effect F1 does not reach the specified F1-Score value, adjust the model parameters and continue model training. The F1-Score value is the harmonic mean of the precision and recall rate, with a maximum of 1 and a minimum of 0.
2. The method for satellite image target recognition based on deep learning according to claim 1, characterized in that: The method for data segmentation and scaling in S2 is as follows: the labeled data in S1 is segmented and saved according to its position information (k1, k2, k3, k4), the segmented data is saved according to its corresponding category label l0, and the data is scaled for input into the model, and it is adjusted to 224*224 according to the size ratio.
3. The method for satellite image target recognition based on deep learning according to claim 1, characterized in that: The method for data division and amplification in S3 is: divide the data set into a training set, a validation set and a test set in a ratio of 7:2:1, the training set is used for model training, the validation set is used to detect whether the model loss continues to decrease, and the test set is used to test the model effect. The training set data is rotated 45 degrees, 90 degrees, and 135 degrees respectively, and the data contrast and brightness are adjusted to amplify the data, and the transformed data is mixed with the original training set to construct a new data set.
4. A satellite image target recognition system based on deep learning, characterized by: The invention comprises a data annotation module (1), a data segmentation and scaling module (2), a data division and amplification module (3), and a model training module (4), wherein the data annotation module (1) is connected to the data segmentation and scaling module (2) via communication, and the data annotation module (1) annotates objects to be identified in the satellite image, wherein the annotation information includes the category of the object, the location of the object, and the easily confused objects, and at the same time transforms the annotation frame from a rectangle to a square, expands the annotation range, and includes the context information around the target to be identified in the annotation range; the data segmentation and scaling module (2) is connected to the data segmentation and amplification module (3) via communication, and the data segmentation and scaling module (2) segments the objects annotated in the satellite image according to their location information, and scales the data size of the data set according to the scale of the convolutional neural network input image; the data segmentation and amplification module (3) divides the data set into a training set, a verification set, and a test set in proportion, and rotates, changes the color of the data in the training set, and amplifies the data quantity; The calculation method of data annotation is as follows: mark the n types of objects to be identified in the satellite image, mark them as 1, 2...n, and also mark the objects similar to the objects to be identified, and mark the similar objects with a unified category mark of 0. When marking, the location information of the target object is also marked. The object is marked by a rectangular box, so the annotation information of each object contains 5 information (l0, l1, l2, l3, l4), l0 is the category information, (l1, l2, l3, l4) is the location information, and the four points correspond to the upper left, lower left, lower right, and upper right respectively. The maximum length L = max (l2-l1, l3-l2) of the object location annotation box is used as the edge, and the annotation box is converted into a square, and then it is enlarged by 10% to intercept more edge and context information and provide more recognition features to the model. Then, its location annotation information is updated, and the new annotation information is (l0, k1, k2, k3, k4); The model training module (4) comprises a training module (401), a recognition module (402), and an evaluation module (403). The training module (401) is connected to the data partitioning and amplification module (3) through communication. The training module (401) inputs the assembled training set and its corresponding label into the network, and trains the model through iterative optimization of the network. When the model training loss no longer decreases, the model is continuously trained using the validation set data. If the model loss no longer decreases, the model is saved. If the loss continues to decrease, the model is continuously iteratively trained using the training set until the loss value is stable, and the model is saved. The training module (401) is connected to the recognition module (402), and the test set data is input into the trained model for recognition. The recognition module (402) is connected to the evaluation module (403), and the recognition result is evaluated. The model in the training module (401) adopts the ResNet-152 network model, and the recognition module (402) adopts the F1-Score. The model training method is as follows: the convolution kernel size of the first convolution is 7*7, and the subsequent convolutions are all 3*3. Except for the first layer, every two layers of the subsequent convolution layers constitute a residual block. The last layer is a fully connected layer for outputting the prediction results. The prediction results are obtained by the sigmoid function to obtain the prediction probability for each category. The network input size is 224*224. The processed training set and its corresponding label are input into the network to train the model, and the validation set data is used to train the model to see whether the model loss function continues to decrease. If it continues to decrease, it indicates that the model has not reached the best. The model is continued to be trained. If it no longer decreases, the model is saved. The model is used to test the test set data and compared with its label results. The result output by the model is the probability that the data is each data category. If the probability of each data category is greater than 0.5, it means that this data is of this category. If the probability of each data category is less than 0.5, it indicates that this data does not contain the object to be identified. The recognition effect is evaluated using F1-Score; The evaluation method of the F1-Score is: F1 is the recognition effect, TP is the correct prediction answer, FP is the wrong prediction of other classes as this class, FN is the prediction of this class label as other class labels, precision is the proportion of positive samples in the positive examples determined by the classifier, and recall is the proportion of positive samples predicted as positive examples in the total positive examples; If the model recognition effect F1 reaches the specified F1-Score value, the model training is completed. If the model recognition effect F1 does not reach the specified F1-Score value, adjust the model parameters and continue model training. The F1-Score value is the harmonic mean of the precision and recall rate, with a maximum of 1 and a minimum of 0.
5. The satellite image target recognition system based on deep learning according to claim 4, characterized in that: The data segmentation and scaling module (2) comprises a data segmentation module (201) and a data scaling module (202), wherein the data segmentation module (201) is connected to the data scaling module (202), and the data segmentation module (201) segments the objects marked in the satellite image according to their position information; the data segmentation module (201) is connected to the data marking module (1), and the data scaling module (202) is connected to the data division and amplification module (3), and the data scaling module (202) scales the data size of the data set according to the scale of the convolutional neural network input image; the data segmentation and scaling method comprises: segmenting and saving the marked data according to its position information (k1, k2, k3, k4), saving the segmented data according to its corresponding category label l0, and scaling the data so as to input the model, and adjusting the data to 224*224 according to the size ratio.
6. The satellite image target recognition system based on deep learning according to claim 4, characterized in that: The data partitioning and amplification module (3) comprises a data partitioning module (301) and a data amplification module (302). The data partitioning module (301) is connected to the data amplification module (302). The data partitioning module (301) is connected to the data scaling module (202) of the data segmentation and scaling module (2). The data partitioning module (301) divides the data set into a training set, a validation set and a test set according to a certain ratio. The data amplification module (302) is connected to the training module (401) of the model training module (4). The data amplification module (302) rotates the data set, changes the color and increases the amount of data. The data partitioning and amplification method comprises: dividing the data set into a training set, a validation set and a test set according to a ratio of 7:2:
1. The training set is used for model training, the validation set is used for detecting whether the model loss continues to decrease, and the test set is used for testing the model effect. The training set data is rotated by 45 degrees, 90 degrees and 135 degrees respectively, and the data contrast and brightness are adjusted. The data is amplified. The transformed data is mixed with the original training set to construct a new data set.
