CN102867196B - Based on the complicated sea remote sensing image Ship Detection of Gist feature learning - Google Patents

Abstract

A ship detection method for complex sea surface remote sensing images based on Gist feature learning, comprising the following steps: step 1, collecting complex sea surface remote sensing image data of different time phases, different sensors, and different scales; step 2, analyzing complex sea surface remote sensing images Block preprocessing, obtaining sample image slices and detecting image slices; step 3, extracting sample image slices and detecting salient features and Gist features of image slices; step 4, training sample image slices according to the salient features and Gist features obtained in step 3, Obtain the training model; step 5, according to the training model obtained in step 4, use the SVM classifier to judge whether the detection image slice contains ships; step 6, find a single ship in the detection image slice based on the improved itti visual attention model. The present invention reduces the false alarm rate as much as possible while ensuring that ships are not missed, can effectively process complex sea surface remote sensing images containing sea clutter, clouds and other interference conditions, and has low computational complexity and strong pertinence.

Description

Ship detection method based on Gist feature learning in complex sea surface remote sensing images

technical field

The invention relates to the technical field of remote sensing image processing, in particular to a ship detection method based on Gist biological visual feature learning for complex sea surface remote sensing images.

Background technique

Ship target detection is a traditional task of coastal countries in the world, and it has a wide range of applications in ship search and rescue, fishing boat monitoring, illegal immigration, territorial defense, anti-drugs, monitoring and management of illegal dumping of oil by ships, etc. . With the development of remote sensing imaging technology, it is possible to use remote sensing images to detect ship targets, and its research objects include the detection of ships themselves and ship wakes. Under complex marine environment conditions, satellite remote sensing images present chaotic fish scales, large reflective areas, and irregularly moving rich textured waves, etc. Small and medium-sized ship targets may be hidden in the complex background clutter, thus affecting the Ship target recognition effect. The sea conditions (wind and waves), weather (clouds and fog), and water color of the ship target image cause unstable sea and land characteristics in the remote sensing image, and there are many interferences such as sea waves, suspended objects on the water surface, and ship wakes, making ship detection difficult.

The existing ship target detection mainly includes the following algorithms: algorithm based on global threshold segmentation, algorithm based on local threshold segmentation, optimal entropy automatic threshold method, algorithm based on distribution model, algorithm based on fractal model, target detection based on feature domain Algorithms, etc., but they generally have the problem of high false alarm rate and poor universality. Among them, the global threshold algorithm cannot automatically adjust the threshold according to the changes in the local area in the image, so the detection results are susceptible to local changes and introduce a large number of false alarms and missed detections. For images with a wide width, due to the imaging mechanism, the sea background has Strong grayscale changes can also lead to false alarms and missed detections. This type of method only uses the gray statistical characteristics of the target in the detection process, without considering the spatial structure information of the target, and the connection between the histogram shape and the image content is also uncertain; When the wind and waves are strong, it is easy to cause a large number of false alarms. At the same time, because this algorithm needs to repeatedly calculate the statistical parameters of the background area in the window, the amount of calculation is huge, and the processing speed is slow, which cannot meet the needs of real-time or near-real-time processing in practical applications; the selection of the threshold value of the optimal entropy automatic threshold method Only the statistical characteristics of the gray level are used, and the spatial structure information of the target is not considered, and the relationship between the distribution of the histogram and the content of the image is also uncertain. In the traditional KSW algorithm, the criterion function is simply defined as the sum of the target gray entropy and the background gray entropy, and the background gray entropy and the target gray entropy occupy the same proportion in the criterion function. This ignores the difference in the proportion of the background and the target in the image, and also ignores the difference in the gray scale range of the background and the target. When the image contains strong sea clutter, good detection results are often not obtained; the algorithm based on the distribution model first needs to make an assumption about the background clutter distribution, which requires certain prior knowledge, but in fact the general situation The background clutter does not strictly obey a certain distribution. Secondly, this type of algorithm needs to count each pixel in the image, so the amount of calculation is large, and it increases with the increase of the sliding window size; the algorithm based on the fractal model believes that the fractal dimension of natural scenery and ship targets has A certain difference is detected according to the difference. However, in actual images, affected by background complexity, random noise, and imaging quality, it is difficult to distinguish natural scenes from artificial targets with a single scale or constant fractal dimension; the target detection algorithm based on feature domains, when the gray distribution of the background is relatively When it is complex, it is greatly affected by noise. At this time, the influence of noise on the feature map will be increased, resulting in mis-segmentation. In addition, in the process of feature conversion, there will be some influence on the outline of the target itself. Using this method to segment the target, the shape information will be lost.

It can be seen that various algorithms are still limited by many conditions, such as the interference of the image background, and the target is affected by weather and illumination changes. Especially for low- and medium-resolution remote sensing data, the ship target appears on the image as For small targets, the probability of missing alarms and false alarms during the monitoring process is high. At the same time, under normal circumstances, visible light images will be disturbed by clouds, oil pollution, sea waves, etc., and it is difficult to establish a background model. Obviously, so it is not easy to separate, especially for black polar ships.

Contents of the invention

In view of the above problems, the present invention proposes a complex sea surface remote sensing image ship detection method based on Gist feature learning.

The technical solution of the present invention is a complex sea surface remote sensing image ship detection method based on Gist feature learning, comprising the following steps:

Step 1, collecting data of complex sea surface remote sensing images;

Step 2, perform block preprocessing on complex sea surface remote sensing images, obtain sample image slices and detect image slices;

Step 3, extract sample image slices and detect salient features and Gist features of image slices;

Step 4, according to the salient features and Gist features obtained in step 3, the sample image slices are trained to obtain the training model;

Step 5, according to the training model obtained in step 4, use the SVM classifier to judge whether any detection image slice contains a ship, if so, enter step 5, otherwise end the detection of the detection image slice;

Step 6, based on the itti visual attention model, find a single ship in the detected image slice, and obtain the ship target.

Moreover, in step 3, the salient feature extraction is performed on any sample image slice and detection image slice, and the implementation method is as follows,

The image slice is decomposed into three feature channels of color, brightness, and direction. The color and brightness feature channels are respectively represented by Gaussian pyramids, and the feature map on the color feature channel and the feature map on the brightness feature channel are generated through the central peripheral difference operation; The direction feature channel uses Gabor filtering to obtain direction pyramids in 4 directions of 0°, 45°, 90°, and 135°, and generates feature maps in each direction through the central peripheral difference operation;

The feature subgraphs of each feature map are fused into color saliency map, brightness saliency map and direction saliency map through local nonlinear fusion, and finally the mean value, standard deviation, local maximum point and local maximum value point of each saliency map are used The distances represent salient features of image slices.

Moreover, in step 2, the Gist feature extraction is performed on any sample image slice and detection image slice, and the implementation method is as follows,

Decompose each image slice into three feature channels of color, brightness, and direction. The color and brightness feature channels are represented by 9-layer Gaussian pyramids respectively, and the feature map on the color feature channel and the feature map on the brightness feature channel are generated by the central peripheral difference operation. Feature map; the direction feature channel uses Gabor filtering to obtain the direction pyramid in 4 directions of 0°, 45°, 90°, and 135°, and obtains the feature map on each direction channel;

Each feature submap of each feature map is divided into 4×4 subgrids, and the average value of 16 subgrids, 4 subgrids in the upper left corner, 4 subgrids in the upper right corner, and 4 subgrids in the lower left corner are respectively calculated , the 4 sub-grids in the upper left corner, the combined mean of the 4 sub-grids in the upper left corner, and the mean of the entire feature submap, and finally use the obtained results to represent the Gist feature vector.

Moreover, in step 6, based on the improved itti visual attention model, searching for a single ship that detects image slices includes the following steps,

Step 6.1, performing primary visual feature extraction on the detected image slice, including brightness, color, direction and texture features;

Step 6.2, saliency map generation, including the four features of brightness, color, direction, and texture obtained in step 6.1 to generate each feature submap through the central peripheral difference operation, and then use the local nonlinear fusion method to fuse to generate a brightness saliency map and a color saliency map , directional saliency map, and texture saliency map, and finally use the local nonlinear fusion method to fuse and generate the overall saliency map; before fused to generate the brightness saliency map, color saliency map, direction saliency map, and texture saliency map, each pixel of the corresponding feature submap The value is squared;

Step 6.3, ship target detection, including detecting a single ship in the image slice by using the return suppression technique on the saliency map obtained in step 6.2, to obtain the ship target.

Moreover, when collecting complex sea surface remote sensing image data, select remote sensing image data from different time phases, different sensors, and different scales. Negative sample images of ships are used as sample image slices.

The method proposed by the present invention first divides the image into blocks, extracts the salient features and Gist features of each sub-image block, trains the sample data according to the extracted feature vector, and then uses the SVM classifier to judge whether the sub-image block contains Gist or not. ship, and finally find a single ship in the sub-image block based on the improved itti visual attention model. The present invention reduces the false alarm rate as much as possible while ensuring that ships are not missed, and can effectively process complex sea surface remote sensing images containing interference from sea clutter, clouds, etc., with low computational complexity and strong pertinence; Complex sea surface remote sensing image data of different phases and different resolutions has good universality; it can quickly process and detect ships with large-format massive remote sensing image data.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Detailed ways

The technical solution of the present invention will be described in detail below in conjunction with the drawings and embodiments.

The ship detection method of complex sea surface remote sensing images based on Gist biological visual feature learning of the present invention is to use the salient features of ships and Gist features to train the sub-image blocks of huge images with SVM classifiers (support vector machine classifiers), A prediction model is obtained, and the support vectors in the model represent the typical characteristics of the ship, and then a single ship is detected for sub-image blocks with suspected ships based on the improved itti visual attention model. Gist features can be found in the existing literature "Drivingme Around the Bend: Learning to Drive from Visual Gist", and Gist features can be calculated using biological visual features. The embodiment process can adopt computer software technology to realize automatic operation, as shown in Figure 1, specifically comprising the following steps:

Step 1, collecting data of complex sea surface remote sensing images.

In specific implementation, the training data coverage should be as much as possible to cope with the subsequent detection steps. The embodiment collects complex sea surface remote sensing image data of different time phases, different sensors, and different scales.

Step 2, perform block preprocessing on complex sea surface remote sensing images, make sample data, obtain sample image slices and detect image slices.

The block processing of the remote sensing image in the embodiment is to block the original huge image into N×N image slices of equal size, that is, each slice includes N×N pixels.

After dividing the collected complex sea surface remote sensing image data of different time phases, different sensors, and different scales into blocks, training samples can be selected from the obtained image slices, and S positive sample image slices (ships) and S negative samples can be manually marked. Image slices (not ships) together constitute the training set of sample image slices. Alternatively, pre-labeled sample image slices can be used directly. All other image slices can be used as detection image slices to be classified. Generally, the entire complex sea surface remote sensing image is taken as the image to be detected, and after the image to be detected is divided into blocks, all sub-image blocks obtained are used as slices of the detected image to be classified.

Wherein, the specific values of N and S can be set by those skilled in the art according to the situation.

Step 3, extract sample image slices and detect salient features and Gist features of the image slices.

In the embodiment, the salient features are extracted from any sample image slice and detection image slice, specifically:

The image slice is decomposed into three feature channels of color, brightness, and direction. The color and brightness feature channels are respectively represented by 9-layer Gaussian pyramids, and the feature maps on the color feature channel and the features on the brightness feature channel are generated by the central peripheral difference operation. Figure: The direction feature channel uses Gabor filtering to obtain a 9-layer direction pyramid in 4 directions of 0°, 45°, 90°, and 135°, and generates a feature map in each direction through the central peripheral difference operation. Among them, the feature map on the color feature channel includes 12 color feature submaps (divided into two groups, one group is the contrast of red and green colors, and the other group is the contrast of yellow and blue colors, each group has 6 pictures), the feature map on the brightness feature channel The graph includes 6 brightness feature submaps (the 9-layer pyramid becomes 6 after the central peripheral difference operation), and the feature map in each direction of the direction feature channel includes 6 directional feature submaps (the 9-layer pyramid undergoes the central peripheral difference operation Then it becomes 6), a total of 24 feature submaps in 4 directions. There are a total of 42 feature submaps, that is, 42=12+6+4×6. Then the feature maps are fused into color, brightness, and direction saliency maps through local nonlinear fusion, in which 12 color feature submaps are fused into 1 color saliency map, 6 luminance feature submaps are fused into 1 brightness saliency map, and 4 A total of 24 direction feature submaps in the direction are fused into one direction saliency map, and a total of 3 saliency maps. Finally, the mean, standard deviation, local maximum point, and distance between local maximum points of each saliency map of the image slice are used, and the 12-dimensional vector composed of 4 values of the 3 saliency maps is used to represent the salience of the image slice. feature. The image mean, standard deviation, local maximum point, specific calculation of the distance between local maximum points, central peripheral difference operation and local nonlinear fusion are existing technologies, which can be found in the literature "L.Itti, C.Koch, E.Niebur, AModel of Saliency-Based Visual Attention for Rapid Scene Analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.20, No.11, pp.1254-1259, Nov 1998.", "L.Itti, C. Koch, Comparison of Feature Combination Strategies for Saliency-Based Visual Attention Systems, In: Proc. SPIE Human Vision and Electronic Imaging IV (HVEI'99), San Jose, CA, Vol.3644, pp.473-82, Bellingham, WA : SPIE Press, Jan 1999.”

In the embodiment, Gist feature extraction is performed on any sample image slice and detection image slice, specifically:

The image slice is decomposed into three feature channels of color, brightness, and direction. The color and brightness feature channels are respectively represented by 9-layer Gaussian pyramids, and the feature maps on the color feature channel and the features on the brightness feature channel are generated by the central peripheral difference operation. Figure: The direction feature channel uses Gabor filtering to obtain 4-layer direction pyramids in 4 directions of 0°, 45°, 90°, and 135°, and takes images of 4 scales of each direction pyramid (that is, the original image is divided by 20, 21 respectively , 22, and 23) as the direction feature submap of the corresponding direction, and the feature maps in the four directions are obtained. The feature map on the color feature channel includes 12 color feature submaps, the feature map on the brightness feature channel includes 6 brightness feature submaps, and the feature map on each direction of the direction feature channel includes 4 direction feature submaps. There are a total of 34 feature submaps, that is, 34=6+12+4×4. Divide each feature submap into 4×4 subgrids, calculate the average value of 16 subgrids respectively, 4 subgrids in the upper left corner, 4 subgrids in the upper right corner, 4 subgrids in the lower left corner, and 4 subnetworks in the upper left corner Grid, the average value of the 4 sub-grids in the upper left corner (that is, calculate the mean value of the 4 sub-grids as a whole) and the mean value of the entire feature sub-graph to obtain a 21-dimensional vector, and finally get a 34×21= 714-dimensional Gist feature vector.

During specific implementation, the number of layers of the Gaussian pyramid can be set by those skilled in the art according to the situation. The more layers there are, the higher the extracted feature dimension will be.

Step 4: Train the sample image slices according to the salient features and Gist features obtained in Step 3 to obtain a training model.

According to the salient features of the sample image slice, and whether it is a positive sample image slice or a negative sample image slice, training can be carried out. The specific training implementation can use the SVM trainer in the prior art, and the training model can be obtained after training all the sample image slices.

Step 5, according to the training model obtained in step 4, use the SVM classifier to judge whether the detected image slice contains ships. For any detected image slice, if the judgment result is contained, go to step 6; if the judged result is not contained, the processing of this detected image slice is ended.

Both the training model and the prediction result are realized by the existing SVM method, for example, the SVM kernel function uses the existing RBF kernel function (radial basis function), which will not be described in detail in the present invention.

Step 6, based on the itti visual attention model, find a single ship in the detected image slice, and obtain the ship target.

The specific implementation of the itti visual attention model is an existing technology. In order to improve the detection accuracy, the embodiment further proposes to improve the itti visual attention model, and based on the improved itti visual attention model, find a single ship for detecting image slices. The improved itti visual attention model adds texture features on the basis of the original itti model, and expands its feature range. At the same time, a square stretching operation is performed before the fusion of feature subgraphs of different scales to increase the salient area and non- Significant area differences, highlighting ship targets.

The steps of the embodiment are as follows:

In step 6.1, primary visual features are extracted from the detected image slices, including brightness, color, orientation and texture features. The original itti visual attention model extracts brightness, color, and direction as primary visual features respectively, and the embodiment increases the extraction of texture features, which can be realized by existing discrete moment transform technology. See the document "VD.Gesu, C.Valent, L .Strinati.Local operators to detect regions of interest[J].Pattern Recognition Letter, 1997,18(11-13):1077-1081."

State of the Art Discrete Moment Transform (DMT) Formula

${\mathrm{<span\; class="notranslate">\; DMT</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\overset{\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\overset{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; q</span>}}$

In the formula, i, j are the row and column values in the detection image slice, g(ir, js) represents the original pixel value of the irth row and js column in the detection image slice; k is the preset window size parameter, for example k=1, The window size is 3×3; r and s are the loop variables in the window, and p and q are indices. Those skilled in the art can set the values of k, r and s according to their needs during specific implementation. In the embodiment, (p=0, q=1), (p=1, q=0), (p=1, q=1) are three DMT texture features, DMT _1.0 DMT _0.1 DMT _1.1 .

Step 6.2, saliency map generation, including the four features of brightness, color, direction, and texture obtained in step 6.1 to generate each feature submap through the central peripheral difference operation, and then use the local nonlinear fusion method to fuse to generate a brightness saliency map and a color saliency map , directional saliency map, and texture saliency map, and finally use the local nonlinear fusion method to fuse and generate the overall saliency map. The value is squared. Due to the square operation, the range of pixel values in the saliency map is stretched, making the layering of the saliency region more obvious and highlighting the ship target.

In the embodiment, the color and brightness feature channels are respectively represented by a 9-layer Gaussian pyramid, and the feature maps on the color feature channel and the feature maps on the brightness feature channel are respectively generated by the central peripheral difference operation; the direction feature channel is obtained by Gabor filtering to obtain 0°, A 9-layer direction pyramid with 4 directions of 45°, 90°, and 135° generates a feature map in each direction through a central peripheral difference operation. Among them, the feature map on the color feature channel includes 12 color feature submaps, the feature map on the brightness feature channel includes 6 brightness feature submaps, and the feature map on each direction of the direction feature channel includes 6 direction feature submaps, A total of 24 feature submaps in 4 directions. There are a total of 42 feature submaps, that is, 42=12+6+4×6. Then, all feature submaps of different scales are used as input, and the feature maps are fused into color, brightness, and direction saliency maps through local nonlinear fusion, in which 12 color feature submaps are fused into 1 color saliency map, and 6 brightness feature submaps The graph is fused into a brightness saliency map, and a total of 24 directional feature submaps in 4 directions are fused into a directional saliency map, and a total of 3 saliency maps.

The feature map on the texture feature channel obtained by discrete moment transform includes 3 texture feature submaps, and the 3 texture feature submaps are fused into 1 texture saliency map, plus the previous 3 feature saliency maps, a total of 4 saliency maps map, and then fuse them into an overall saliency map, which integrates all features.

Step 6.3, ship target detection, the saliency map obtained in step 6.2 is used to detect a single ship target in the image slice through the return suppression technique. Existing return suppression technology can be found in "L.Itti, C.Koch, E.Niebur, A Model of Saliency-Based Visual Attention for Rapid Scene Analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.20, No.11, pp. 1254-1259, Nov 1998." The present invention will not go into details.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (2)

1. a complex sea surface remote sensing image ship detection method based on Gist feature learning, is characterized in that, comprises the following steps: Step 1, collecting data of complex sea surface remote sensing images; Step 2, perform block preprocessing on complex sea surface remote sensing images, obtain sample image slices and detect image slices; Step 3, extract any sample image slice and detect the salient features and Gist features of the image slice; In step 3, perform salient feature extraction on any sample image slice and detection image slice, and the implementation method is as follows, Decompose any sample image slice or detection image slice into three feature channels of color, brightness, and direction. The color and brightness feature channels are represented by Gaussian pyramids respectively, and the feature map and brightness on the color feature channel are generated through the central peripheral difference operation The feature map on the feature channel; the direction feature channel uses Gabor filtering to obtain the direction pyramid in 4 directions of 0°, 45°, 90°, and 135°, and the feature map in each direction is generated by the central peripheral difference operation; The feature subgraphs of each feature map are fused into color saliency map, brightness saliency map and direction saliency map through local nonlinear fusion, and finally the mean value, standard deviation, local maximum point and local maximum value point of each saliency map are used The distance represents the salient feature of any sample image slice or detection image slice; In step 3, the Gist feature extraction is performed on any sample image slice and detection image slice, and the implementation method is as follows, Decompose any sample image slice or detection image slice into three feature channels of color, brightness, and direction. The color and brightness feature channels are represented by 9-layer Gaussian pyramids respectively, and the feature maps on the color feature channel and The feature map on the brightness feature channel; the direction feature channel uses Gabor filtering to obtain the direction pyramid in 4 directions of 0°, 45°, 90°, and 135°, and obtains the feature map on each direction channel; Each feature submap of each feature map is divided into 4×4 subgrids, and the average value of 16 subgrids, 4 subgrids in the upper left corner, 4 subgrids in the upper right corner, and 4 subgrids in the lower left corner are respectively calculated , the average value of the four subnetworks in the lower right corner after merging and the average value of the entire feature submap, and finally use the obtained results to represent the Gist feature; Step 4, according to the salient features and Gist features obtained in step 3, the sample image slices are trained to obtain the training model; Step 5, according to the training model obtained in step 4, use the SVM classifier to judge whether any detection image slice contains a ship, if so, enter step 5, otherwise end the detection of the detection image slice; Step 6, based on the itti visual attention model, look for a single ship that detects the image slice, and obtain the ship target; including the following steps, Step 6.1, performing primary visual feature extraction on the detected image slice, including brightness, color, direction and texture features; Step 6.2, saliency map generation, including the four features of brightness, color, direction, and texture obtained in step 6.1 to generate each feature submap through the central peripheral difference operation, and then use the local nonlinear fusion method to fuse to generate a brightness saliency map and a color saliency map , directional saliency map, and texture saliency map, and finally use the local nonlinear fusion method to fuse and generate the overall saliency map; before fused to generate the brightness saliency map, color saliency map, direction saliency map, and texture saliency map, each pixel of the corresponding feature submap The value is squared; Step 6.3, ship target detection, including detecting a single ship in the image slice by using the return suppression technique on the saliency map obtained in step 6.2, to obtain the ship target. 2. according to the complex sea surface remote sensing image ship detection method based on Gist feature learning of claim 1, it is characterized in that: when collecting the data of complex sea surface remote sensing image, select remote sensing image data from different time phases, different sensors, and different scales , after preprocessing the complex sea surface remote sensing image into blocks, mark the positive sample images containing ships and the negative sample images without ships as sample image slices.