CN114220019B - Lightweight hourglass type remote sensing image target detection method and system - Google Patents
Lightweight hourglass type remote sensing image target detection method and system Download PDFInfo
- Publication number
- CN114220019B CN114220019B CN202111323948.7A CN202111323948A CN114220019B CN 114220019 B CN114220019 B CN 114220019B CN 202111323948 A CN202111323948 A CN 202111323948A CN 114220019 B CN114220019 B CN 114220019B
- Authority
- CN
- China
- Prior art keywords
- target detection
- network
- remote sensing
- hourglass
- sensing image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 91
- 238000000034 method Methods 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 19
- 238000012360 testing method Methods 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000012795 verification Methods 0.000 claims abstract description 7
- 230000000644 propagated effect Effects 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 9
- 238000010606 normalization Methods 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 230000006835 compression Effects 0.000 claims description 2
- 238000007906 compression Methods 0.000 claims description 2
- 230000008034 disappearance Effects 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 230000001629 suppression Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 101100494773 Caenorhabditis elegans ctl-2 gene Proteins 0.000 description 1
- 101100112369 Fasciola hepatica Cat-1 gene Proteins 0.000 description 1
- 101100005271 Neurospora crassa (strain ATCC 24698 / 74-OR23-1A / CBS 708.71 / DSM 1257 / FGSC 987) cat-1 gene Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a lightweight hourglass type remote sensing image target detection method and a lightweight hourglass type remote sensing image target detection system, wherein the method comprises the steps of preprocessing an acquired remote sensing image data set, and dividing the acquired remote sensing image data set into a training data set, a verification data set and a test data set; constructing a target detection network model, wherein the target detection network model comprises a lightweight hourglass type network and a characteristic pyramid network; inputting a training dataset remote sensing image, extracting features by using a lightweight hourglass type network, and inputting the extracted features into a feature pyramid network to finish feature coding; obtaining a plurality of candidate frames aiming at the coding result, and selecting an optimal candidate frame as output to obtain a predicted value; the error of the predicted value and the true value is reversely propagated, and training of the target detection network model is completed; and finishing prediction and classification on the test data set by using the trained target detection network model. The invention has high detection precision on small targets and higher convergence rate of training.
Description
Technical Field
The invention relates to the field of remote sensing image processing, in particular to a lightweight hourglass type remote sensing image target detection method and system.
Background
The target detection of the remote sensing image is a technical means for detecting an interested region or object on a ground object image shot by a high-resolution satellite, and the interested object can be selected in a rectangular frame form, and the category confidence of the object is given. Objects to be detected often include small objects such as airplanes, ships, courts, vehicles, and larger objects such as bridges, railway stations, ports, and the like. Along with the improvement of the resolution of the remote sensing image, the information contained in the image is more and more abundant, and the target detection technology of the remote sensing image becomes an important ring in the field of remote sensing image analysis, and has great significance in natural disaster assessment, resource exploration, military research and the like.
At present, the target detection algorithm based on deep learning can be mainly divided into two major categories, namely a two-stage target detection algorithm and a single-stage target detection algorithm. The two-stage target detection algorithm mainly comprises RCNN, fast-RCNN, R-FCN and the like, and the core principle is that a series of candidate frames of samples are generated by the algorithm, and then objects in the candidate frames are classified by a convolutional neural network. The single-stage target detection algorithm mainly comprises YOLO, YOLO9000, YOLOv3, SSD, retinaNet and the like, and the core principle is that the problem of positioning and classifying a target frame is directly converted into a regression problem, and the positioning and classifying of an object are obtained by solving the regression problem. Generally, the two-stage target detection algorithm can achieve higher precision, but has lower calculation efficiency; in contrast, single-stage target detection algorithms can maintain higher computational efficiency, but suffer from a slight drop in detection accuracy.
In the object detection algorithm based on deep learning, in order to acquire a feature map with higher-level semantic features, a large number of convolution operations accompanied by downsampling are often used, and the resolution of the feature map is inevitably reduced while rich semantic features are acquired, so that the detection performance of a small object is easily reduced. Aiming at the problem, the common method is to carry out cascading or fusion on the feature images of different levels, so as to ensure that the feature images have better semantic and fine-grained information. If YOLOv3 builds a feature pyramid after the feature extraction backbone network, the high-level features are cascaded with the low-level features through upsampling, and network prediction is output at multiple levels, so that the network has the characteristic of multi-scale target detection.
In the target detection task of the remote sensing image, the image to be processed is often larger in resolution, a large number of small targets exist, each pixel point possibly contains important information, and the common cascade or fusion method of different-level feature images is extremely limited in improvement. How to improve the detection precision of small targets in target detection while maintaining the calculation efficiency, and fully utilize the context information to assist the target detection task becomes a considerable problem.
Disclosure of Invention
Aiming at the defect that the existing target detection algorithm cannot realize better detection performance of a small target in a remote sensing image, the invention provides a lightweight hourglass type remote sensing image target detection method and system.
The technical scheme adopted by the invention is as follows:
a lightweight hourglass type remote sensing image target detection method comprises the following steps of
Preprocessing an acquired remote sensing image data set, and dividing the acquired remote sensing image data set into a training data set, a verification data set and a test data set;
constructing a target detection network model, wherein the target detection network model comprises a lightweight hourglass type network and a characteristic pyramid network;
inputting a training dataset remote sensing image, extracting features by using a lightweight hourglass type network, and inputting the extracted features into a feature pyramid network to finish feature coding;
obtaining a plurality of candidate frames aiming at the coding result, and selecting an optimal candidate frame as output to obtain a predicted value;
the error of the predicted value and the true value is reversely propagated, and training of the target detection network model is completed;
and finishing prediction and classification on the test data set by using the trained target detection network model.
Further, the preprocessing includes converting a detection frame in the remote sensing image dataset from a form of "center point+width and height" to a form of "upper left corner+lower right corner".
Further, the lightweight hourglass network is formed into various levels from a plurality of stacked hourglass base modules.
Further, the construction method of the hourglass-shaped basic module is as follows:
for inputConvolution and corresponding batch normalization and activation functions were performed using a convolution kernel of 3*3, step 1 and 3*3, respectively, step 2Processing to form two branches, the corresponding output results are +.>Wherein h is 1 ＝2h 2 ，w 1 ＝2w 2 ，c 1 ＝2c 2 ；
For H 1 Performing double nearest neighbor interpolation to obtain an up-sampling resultH 2 Width and height of H 1 Keep consistent due to I to H 2 The width and height of the feature undergo compression and expansion in sequence, and the change process of the feature imitates the shape of an hourglass, so that the feature is called an hourglass structure;
in the channel dimension pair S 1 And H 2 Cascading results are thatWherein c 3 ＝c 1 +c 2 ，S 2 Fine granularity information and semantic information before and after convolution are contained;
for cascade result S 2 The number of channels is compressed to c using a 1*1, step-1 convolution kernel 1 To reduce the parameter calculation amount, output the result asS 3 The size of the (C) is consistent with the size of the I, and subsequent residual connection is carried out;
pair S using channel attention mechanism 3 The channel weight of the model is adjusted to make the model more concentrated in high-value characteristic information, and the re-weighting result is that
In the form of pixel addition to S 4 Residual connection is carried out with I, so that the problem of gradient disappearance caused by deep network is prevented, and the output result of the module is that
And stacking the hourglass type module structures according to the number of basic modules set by each level of the lightweight hourglass type network to form a level of a backbone network, wherein the input and output sizes of the modules of the same level are consistent.
Further, the construction method of the lightweight hourglass network is as follows:
stacking the hourglass-shaped base modules to form a hierarchy of the network;
and adding a 3*3 convolution kernel with the step length of 2 between different levels to carry out convolution kernel and corresponding batch normalization and activation function processing to adjust the number of channels and acquire higher-level semantic information, and connecting the levels in series to form a lightweight hourglass type target detection backbone network.
Further, the lightweight hourglass network comprises five levels, wherein the five levels respectively comprise 1, 2, 4 and 2 stacked basic modules, the feature width and height of the later level are 1/2 of that of the former level, and the number of channels is 2 times that of the former level.
Further, the last three levels of output features of the lightweight hourglass network are input into the feature pyramid network, and the downstream features are up-sampled and then cascaded with the upstream features.
Further, the loss of the width and the height of the prediction frame is calculated by adopting the Euclidean distance loss function, and the loss of the center point, the confidence and the category of the prediction frame is calculated by adopting the cross entropy loss function.
Further, a non-maximum suppression algorithm is adopted to select the optimal candidate frame for output.
A system for a light-weight hourglass type remote sensing image target detection method comprises the following steps:
the remote sensing image data acquisition unit is used for preprocessing an acquired remote sensing image data set and dividing the remote sensing image data set into a training data set, a verification data set and a test data set;
the method comprises the steps of constructing a target detection network model unit, wherein the target detection network model unit is used for constructing a target detection network model, and the target detection network model comprises a lightweight hourglass network and a characteristic pyramid network;
the training target detection network model unit is used for initializing convolution kernel weights and offsets of all layers of the network by using Gaussian distribution with zero mean value, and performing optimization iteration on the network model by adopting the back propagation of an Adam optimizer;
and the target detection unit is used for completing a target detection task on the remote sensing images in the test set.
The invention has the beneficial effects that:
(1) The invention adopts a lightweight hourglass type module to replace the traditional residual error module, and combines the characteristic diagrams of different convolutions to ensure that the output of the basic module has both semantic information and fine granularity information;
(2) Taking the YOLOv3 target detection algorithm in the prior art as an example, the method has five levels, and each level comprises 1, 2, 8 and 4 basic modules respectively; each level of the target detection algorithm provided by the invention comprises 1, 2, 4 and 2 basic modules respectively, so that higher calculation efficiency can be effectively maintained;
(3) The output of each basic module has better semantic information and fine granularity information, and compared with YOLOv3, the method has higher detection precision on small targets and higher convergence rate of training.
Drawings
FIG. 1 is a flow chart of an implementation of a lightweight hourglass type remote sensing image target detection method.
Fig. 2 is a first level of a lightweight hourglass type object detection model, comprising 1 base module.
FIG. 3 is a backbone network structure of the object detection model, each hierarchy including a different number of basic modules.
Fig. 4 (a) to fig. 4 (b) are comparisons between the YOLOv3 method and the average accuracy of the DIOR dataset when the target detection task is completed by the method according to the present invention, wherein fig. 4 (a) is the average accuracy of the YOLOv3 detection, and fig. 4 (b) is the average accuracy of the detection according to the method according to the present invention.
Fig. 5 (a) to 5 (b) are effects of the YOLOv3 method and the proposed method on the same remote sensing image, wherein fig. 5 (a) is a detection result of YOLOv3, and fig. 5 (b) is a detection result of the proposed method.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but embodiments of the present invention are not limited thereto.
Examples
As shown in fig. 1, a light-weight hourglass type remote sensing image target detection method is taken as an example of a remote sensing image data set with a 23463 Zhang Fenbian rate of 800×800×3, and includes the following steps:
s1, preprocessing an acquired remote sensing image data set, and dividing the acquired remote sensing image data set into a training data set, a verification data set and a test data set;
the preprocessing in this embodiment refers to converting the data labeling mode from the form of "center point+width and height" to the form of "upper left corner+lower right corner", dividing the data set into a training data set, a verification data set and a test data set according to the ratio of 4:1:5, and adjusting the size of the input image to be a fixed value.
The size of the input image is adjusted to 416×416×3.
S2, constructing a target detection network model, wherein the target detection network model comprises a lightweight hourglass network and a characteristic pyramid network, randomly initializing the weight and bias of a convolution kernel by using Gaussian distribution with the mean value of zero, and inputting a divided training data set into the lightweight hourglass network for learning;
the lightweight hourglass network serves as a backbone network and the feature pyramid network serves as a neck network.
The proposed backbone network with the lightweight hourglass module as the base module was applied to a conventional YOLOv3 target detection model, with batch normalization and linear rectification function activation after each convolution.
The lightweight hourglass network is formed into various levels by a plurality of hourglass-type basic modules which are stacked.
The specific constitution process of the hourglass type basic module comprises the following steps:
further, the specific process of the step 2 is as follows:
step 2.1, input I ε R 416×416×3 Firstly, carrying out convolution operation with convolution kernels of 32 3*3 to expand channel dimension, and obtaining output characteristic I after batch normalization and linear array rectification activation function 1 ∈R 416×416×32 ；
Step 2.2, I 1 Input to a first level of a lightweight hourglass type object detection model, the structure of which is shown in FIG. 2:
a first layer: convolutional layer Conv1, input I 1 Convolving with 64 convolution kernels of 3*3 with step length of 2, and performing batch normalization and linear rectification activation processing to obtain output characteristic I 2 ∈R 208×208×64 ；
A second layer: two convolution branches Conv2 and Conv3 are input into the output of the upper layer, are respectively subjected to convolution operation with 64 convolution kernels with the step length of 1 and 32 convolution kernels with the step length of 2, and are subjected to batch normalization and linear rectification activation processing to respectively obtain output characteristics S 1 ∈R 208×208×64 And H 1 ∈R 104×104×32 ；
Third layer: up sample layer UpSample1, input output H of the previous layer 1 Up-sampling result H is obtained by using double nearest neighbor interpolation 2 ∈R 208×208×32 ，H 2 Width and height of H 1 Keeping consistency so as to facilitate subsequent cascading operation;
fourth layer: cascade Cat1, input is output S of the second layer 1 And output H of the third layer 2 The two are cascaded from the channel dimension to obtain an output characteristic S 2 ∈R 208×208×96 ；
Fifth layer: the output of the previous layer is input into the Conv4 of the convolution layer, the convolution operation is carried out with 64 3*3 convolution kernels with the step length of 1, and the output characteristic S is obtained through batch normalization and linear rectification activation processing 3 ∈R 208×208×64 ，S 3 The size of the residual error is consistent with the size of the I so as to ensure that the subsequent residual error connection is smoothly carried out;
sixth layer: attention layer Atten1 is input to the output of the previous layer, and S is obtained through two-dimensional self-adaptive mean value pooling and one-dimensional convolution 3 In the channelWeight matrix W in dimension 1 And then it is combined with S 3 Multiplying to obtain an output S after adjusting the channel weight 4 ∈R 208 ×208×64 ；
Seventh layer: residual connection layer Res1, input the output of the previous layer, and output the residual connection layer Res1 and I 2 Residual connection is carried out in a mode of adding corresponding pixel points, and the output O epsilon R of the current module is obtained 208×208×64 ；
Step 2.3, the lightweight hourglass type target detection model totally comprises 5 layers, each layer comprises 1, 2, 4 and 2 basic modules, and the constructed basic modules are repeatedly stacked according to the number of the modules of the corresponding layers. Taking level 1 as an example, the level has 1 basic module, as shown in fig. 2;
and 2.4, directly connecting the basic modules of the same level to each other in the model of the step 2.3, convolving the basic modules of different levels by adopting a convolution kernel with 3*3 and a step length of 2, wherein the characteristic width and height of the later level are 1/2 of that of the former level, and the number of channels is 2 times that of the former level. Then in this example the first level output is L 1 ∈R 208×208×64 The second level output is L 2 ∈R 104 ×104×128 The third level output is L 3 ∈R 52×52×256 The fourth level output is L 4 ∈R 26×26×512 The fifth level output is L 5 ∈R 13×13×1024 The specific structure is shown in figure 3;
and 2.5, taking the output of the main network as the input of the characteristic pyramid network, transmitting the input of the main network into the neck network, finishing the coding of the characteristics, and obtaining the predicted output of the network through subsequent decoding.
S3, calculating the loss of the width and the height of the prediction frame by using an Euclidean distance loss function, and calculating the loss of the center point, the confidence coefficient and the category of the prediction frame by using a cross entropy loss function, wherein the summation of the four losses is used as the prediction loss of the whole network model;
s4, carrying out back propagation on errors by using an Adam optimizer, carrying out iterative updating on the weight and bias of the convolution kernel, and when the loss function reaches a minimum value or reaches a set maximum iterative step number, regarding the loss function as obtaining a current optimal depth network model;
and S5, adjusting the test data set to 416 x 3, and inputting the test data set into the network model obtained in the step 4 to obtain a target detection prediction result of the test set.
In this embodiment, the DIOR dataset is used to compare YOLOv3 with the lightweight hourglass type remote sensing image target detection method provided by the invention, after training for 50 rounds, the average precision is used as an evaluation index, and the comparison result is shown in table 1.
The comparison result is visualized in fig. 4 (a) and fig. 4 (b), wherein fig. 4 (a) shows the detection effect of YOLOv3 on each category, the average detection precision is 63.91%, and fig. 4 (b) shows the detection effect of the method provided by the invention on each category, and the average detection precision is 66.60%. By comparison, for most of targets in the used data set, compared with YOLOv3, the method provided by the invention can obtain better detection effect, and the detection performance of small targets is further improved.
TABLE 1 Experimental results for target detection on DIOR datasets
Fig. 5 (a) to 5 (b) show the target detection contrast effect of YOLOv3 and the proposed method of the present invention on the same remote sensing image, wherein fig. 5 (a) is the detection result of YOLOv3, and fig. 5 (b) is the detection result of the proposed method of the present invention. Compared with YOLOv3, the detection frame prediction method provided by the invention has higher confidence coefficient, and the boundary of the detection frame is more attached to an actual target.
The embodiments described above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the embodiments described above, and any other changes, modifications, substitutions, combinations, and simplifications that do not depart from the spirit and principles of the present invention should be made in the equivalent manner, and are included in the scope of the present invention.
Claims (6)
1. A light-weight hourglass type remote sensing image target detection method is characterized by comprising the following steps of
Preprocessing an acquired remote sensing image data set, and dividing the acquired remote sensing image data set into a training data set, a verification data set and a test data set;
constructing a target detection network model, wherein the target detection network model comprises a lightweight hourglass type network and a characteristic pyramid network;
inputting a training dataset remote sensing image, extracting features by using a lightweight hourglass type network, and inputting the extracted features into a feature pyramid network to finish feature coding;
obtaining a plurality of candidate frames aiming at the coding result, and selecting an optimal candidate frame as output to obtain a predicted value;
the error of the predicted value and the true value is reversely propagated, and training of the target detection network model is completed;
the test data set is predicted and classified by using the trained target detection network model;
the lightweight hourglass network comprises a plurality of hourglass-shaped basic modules which are stacked to form each level;
the construction method of the hourglass-shaped basic module comprises the following steps:
for inputConvolution and corresponding batch normalization and activation function processing are carried out on convolution kernels with the step length of 2 by using 3*3 and the step length of 1 and 3*3 respectively to form two branches, and corresponding output results are +.>Wherein h is 1 ＝2h 2 ，w 1 ＝2w 2 ，c 1 ＝2c 2 ；
For H 1 Performing double nearest neighbor interpolation to obtain an up-sampling resultH 2 Width and height of H 1 Keep consistent due to I to H 2 The width and height of the feature undergo compression and expansion in sequence, and the change process of the feature imitates the shape of an hourglass, so that the feature is called an hourglass structure;
in the channel dimension pair S 1 And H 2 Cascading results are thatWherein c 3 ＝c 1 +c 2 ，S 2 Fine granularity information and semantic information before and after convolution are contained;
for cascade result S 2 The number of channels is compressed to c using a 1*1, step-1 convolution kernel 1 To reduce the parameter calculation amount, output the result asS 3 The size of the (C) is consistent with the size of the I, and subsequent residual connection is carried out;
pair S using channel attention mechanism 3 The channel weight of the model is adjusted to make the model more concentrated in high-value characteristic information, and the re-weighting result is that
In the form of pixel addition to S 4 Residual connection is carried out with I, so that the problem of gradient disappearance caused by deep network is prevented, and the output result of the module is that
Stacking the hourglass module structures according to the number of basic modules set by each level of the lightweight hourglass network to form a level of a backbone network, wherein the input and output sizes of the modules of the same level are kept consistent;
the construction method of the lightweight hourglass network comprises the following steps:
stacking the hourglass-shaped base modules to form a hierarchy of the network;
a 3*3 convolution kernel with the step length of 2 is added between different levels to carry out convolution kernel and corresponding batch normalization and activation function processing so as to adjust the number of channels and acquire higher-level semantic information, and the levels are connected in series to form a lightweight hourglass type target detection backbone network;
the lightweight hourglass network comprises five levels, wherein the five levels respectively comprise 1, 2, 4 and 2 stacked basic modules, the feature width and height of the later level are 1/2 of that of the former level, and the number of channels is 2 times that of the former level.
2. The method of claim 1, wherein the preprocessing includes converting a detection frame in the remote sensing image dataset from a "center point + width-height" form to a "top left corner + bottom right corner" form.
3. The method for detecting the target of the lightweight hourglass type remote sensing image according to claim 1, wherein the last three levels of the lightweight hourglass type network output features are input into a feature pyramid network, and are cascaded with the uplink features after the downlink features are up-sampled.
4. The method for detecting the light-weight hourglass type remote sensing image target according to claim 1, wherein the loss of the width and the height of the prediction frame is calculated by using a Euclidean distance loss function, and the loss of the central point, the confidence and the category of the prediction frame is calculated by using a cross entropy function.
5. The method for detecting the light-weight hourglass type remote sensing image target according to claim 1, wherein a non-maximum suppression algorithm is adopted to select the optimal candidate frame for output.
6. A system for implementing the lightweight hourglass type remote sensing image object detection method of any one of claims 1-5, comprising:
the remote sensing image data acquisition unit is used for preprocessing an acquired remote sensing image data set and dividing the remote sensing image data set into a training data set, a verification data set and a test data set;
the method comprises the steps of constructing a target detection network model unit, wherein the target detection network model unit is used for constructing a target detection network model, and the target detection network model comprises a lightweight hourglass network and a characteristic pyramid network;
the training target detection network model unit is used for initializing convolution kernel weights and offsets of all layers of the network by using Gaussian distribution with zero mean value, and performing optimization iteration on the network model by adopting the back propagation of an Adam optimizer;
and the target detection unit is used for completing a target detection task on the remote sensing images in the test set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111323948.7A CN114220019B (en) | 2021-11-10 | 2021-11-10 | Lightweight hourglass type remote sensing image target detection method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111323948.7A CN114220019B (en) | 2021-11-10 | 2021-11-10 | Lightweight hourglass type remote sensing image target detection method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114220019A CN114220019A (en) | 2022-03-22 |
| CN114220019B true CN114220019B (en) | 2024-03-29 |
Family
ID=80696796
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111323948.7A Active CN114220019B (en) | 2021-11-10 | 2021-11-10 | Lightweight hourglass type remote sensing image target detection method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114220019B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117542104B (en) * | 2024-01-09 | 2024-04-30 | 浙江图讯科技股份有限公司 | Face three-dimensional key point detection method based on self-supervision auxiliary learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532859A (en) * | 2019-07-18 | 2019-12-03 | 西安电子科技大学 | Remote Sensing Target detection method based on depth evolution beta pruning convolution net |
| CN111723748A (en) * | 2020-06-22 | 2020-09-29 | 电子科技大学 | Infrared remote sensing image ship detection method |
| CN111753677A (en) * | 2020-06-10 | 2020-10-09 | 杭州电子科技大学 | Multi-angle remote sensing ship image target detection method based on characteristic pyramid structure |
-
2021
- 2021-11-10 CN CN202111323948.7A patent/CN114220019B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532859A (en) * | 2019-07-18 | 2019-12-03 | 西安电子科技大学 | Remote Sensing Target detection method based on depth evolution beta pruning convolution net |
| CN111753677A (en) * | 2020-06-10 | 2020-10-09 | 杭州电子科技大学 | Multi-angle remote sensing ship image target detection method based on characteristic pyramid structure |
| CN111723748A (en) * | 2020-06-22 | 2020-09-29 | 电子科技大学 | Infrared remote sensing image ship detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114220019A (en) | 2022-03-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111612066B (en) | Remote sensing image classification method based on depth fusion convolutional neural network | |
| CN108960143B (en) | Ship detection deep learning method in high-resolution visible light remote sensing image | |
| CN112991354B (en) | High-resolution remote sensing image semantic segmentation method based on deep learning | |
| CN112347888B (en) | Remote sensing image scene classification method based on bi-directional feature iterative fusion | |
| He et al. | A dual global–local attention network for hyperspectral band selection | |
| CN108288270B (en) | Target detection method based on channel pruning and full convolution deep learning | |
| CN110969087B (en) | Gait recognition method and system | |
| CN113674334B (en) | Texture recognition method based on depth self-attention network and local feature coding | |
| CN113240683B (en) | Attention mechanism-based lightweight semantic segmentation model construction method | |
| CN112381733B (en) | Image recovery-oriented multi-scale neural network structure searching method and network application | |
| CN111898461B (en) | Time sequence behavior segment generation method | |
| CN111626267B (en) | Hyperspectral remote sensing image classification method using void convolution | |
| CN109002771B (en) | Remote sensing image classification method based on recurrent neural network | |
| CN114220019B (en) | Lightweight hourglass type remote sensing image target detection method and system | |
| CN112926533A (en) | Optical remote sensing image ground feature classification method and system based on bidirectional feature fusion | |
| CN115393690A (en) | Light neural network air-to-ground observation multi-target identification method | |
| CN115457332A (en) | Image multi-label classification method based on graph convolution neural network and class activation mapping | |
| CN116824133A (en) | Intelligent interpretation method for remote sensing image | |
| CN115953902A (en) | Traffic flow prediction method based on multi-view space-time diagram convolution network | |
| CN115908907A (en) | Hyperspectral remote sensing image classification method and system | |
| CN114494284A (en) | Scene analysis model and method based on explicit supervision area relation | |
| CN110866560A (en) | Symmetric low-rank representation subspace clustering method based on structural constraint | |
| CN117315324B (en) | Lightweight class detection method and system for Mars rugged topography | |
| CN113378731B (en) | Green space water system vector extraction method based on convolutional neural network and energy constraint | |
| CN111339825B (en) | Model training method based on characteristic relation atlas learning and data classification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |