Intelligent re-recognition algorithm for specific ship target in busy waters under the actual scene

Abstract

The traditional re-recognition algorithm needs to find or design the characteristics with better robustness to light, scale, and deformation. The quality of the feature directly affects the recognition performance and the uncertainty is high. In addition, it needs supervision and training, and has the higher training time and space complexity. To address this problem, a new intelligent re-recognition algorithm for specific ship target in busy waters under the actual scene is proposed in this paper. Combining the existing feature extraction model and graph model, the graph structure is used to describe the identity relationship between the samples. Two points with side connections have the same identity label. Then the multi-layer graph structure is built. After obtaining the block of the divided area, the similarity between the two samples of the link is calculated and the weight of the edge is obtained. Labeled samples are built according to the selected initial area. The energy loss of the graph model is obtained by estimating the pixel likelihood energy function with different labels of pixels and areas. A graph structure is obtained by minimizing the energy loss, which is the intelligent recognition result of specific ship target. For the large-scale data, the problem of incremental processing is solved by incremental maintenance. Experimental results show that the proposed algorithm has high recognition precision.

Keywords

Actual scene busy waters specific ship target intelligence re-recognition

1 Introduction

In the visual sensor network without overlapping vision, the recognition of a recognized ship target again for the change of conditions (monitoring scene, illumination condition, and target pose) refers to target re-recognition [1]. Target re-recognition technology has important research significance in the fields of intelligent monitoring and multi-target tracking. In recent years, many scholars at home and abroad have received extensive attention and research [2]. In the actual scene, the traffic is busy, and the accident is unavoidable. The effective re-recognition of specific ship target in busy waters has a great effect on economic development [3]. The ship target intelligent re-recognition technology is also of great significance for accurate tracking of ship target and accurate control.

The re-recognition of ship target is facing great challenges. Because of the influence of illumination change, shooting angle, posture change, occlusion, and camera’s own characteristics, the appearance characteristics of the same ship target for different cameras have changed a lot. Therefore, how to establish a robust ship target re-recognition algorithm is one of the hotspots of research at present.

The ship target re-recognition algorithm is mainly divided into two kinds of algorithms, which are based on feature learning and based on metric learning, respectively. The algorithm based on feature learning selects or designs an appearance description feature, which makes the inner-class similarity high and the between-class discrimination strong. In the literature [4], 8 color features and 21 texture features are used to detect the characteristics of different attribute features of the ship. Then the posteriori probability of each attribute is used to form a middle layer feature and ship re-recognition. In the literature [5], the ship target are divided into multiple blocks and the color feature and dense SIFT feature of each block are extracted. And then unsupervised learning is used to obtain the saliency characteristics of the ship target for recognition. In the literature [6], in the process of recognition, weights of all kinds of feature are dynamically adjusted to select the most important feature for ship target re-recognition by using the color characteristics and texture features. In the literature [7], the color and structure information of the ship target is used to research the feature construction and measure selection of the re-recognition problem. Although the above feature learning algorithms have achieved good research results, such algorithms need to find or design the features with better robustness to illumination, scale, and deformation factors, and the quality of the algorithm directly affects the recognition performance.

The algorithm based on metric learning is to obtain a model by leaning from image feature space, which makes the feature distance from the same category closer, while the distance from different categories farther, so as to maximize the matching accuracy. In the literature [8], a distance metric mechanism for large distance nearest neighbor is proposed. It can minimize the distance between the correctly matched targets and maximize the distance between the mismatched targets. In the literature [9], the proposed relative distance comparison probability model makes the maximum distance between the correctly matched targets less than the mismatched targets. In the literature [10], ship target recognition problem is converted into a retrieval sorting problem. A sorted support vector machine is proposed. High recognition performance is obtained by combining the local Fisher feature and the principal component analysis dimension reduction technique. But the metric learning algorithm needs to be supervised and trained, the training time and the space complexity are high, and the distance measure affects the recognition effect of the algorithm.

Based on the above analysis, a new intelligent re-recognition algorithm for specific ship target in busy waters under the actual scene is proposed in this paper. The structure of this paper is as follows.

Section 1: The significance of intelligent re-recognition algorithm for specific ship target in busy waters under the actual scene, and the current re-recognition algorithms and the disadvantages are analyzed.

Section 2: The proposed intelligent re-recognition algorithm is given.

Section 3: The recognition accuracy of the proposed algorithm is verified by experiment.

Section 4: Discussions of the proposed algorithm.

Section 5: Conclusions and the further research.

2 Material and methods

2.1 Graph model building and region segmentation

Combined with the existing feature extraction model and graph model [11], the graph structure is used to describe the identity relationship between samples. The point represents the sample, the 2 points with the side connection have the same identity label, and the multi-layer graph structure is built, as shown in Fig. 1.

Fig. 1

Multi-layer graph structure.

One of the most important steps to build multi-layer graph structure is to obtain the partition. The areas with stable features and rich information are focused on. In this way, the extracted invariant feature is more robust.

The region segmentation of the target is completed by the color segmentation operator and the shape segmentation operator.

The color segmentation operator is defined as

$C (j, σ) = \frac{1}{χ} \sum_{v = 1}^{σ} \sum^{} d (p_{i, j - v}, p_{i, j + v})$ (1) where p_i,j-v ∈ I, p_i,j+v ∈ I, I is the foreground image of busy waters in actual scene, p_i,j is the pixel with the global coordinates (i, j), d (p_i,j-v, p_i,j+v) is the Euclidean distance of two pixels in the RUB space, χ is the total number of accumulation, σ is the fixed value, which determines the image region used to calculate the color segmentation operator, C (j, σ) is the difference of color feature between the upper and lower σ rows with division line of the jth row. On this basis, color segmentation operator is used to calculate the color segmentation line c between different regions. The number of areas that need to be divided is determined according to the ship target.

The shape segmentation operator is defined as $S (j, σ) = \frac{1}{κ} \sum_{v = 1}^{κ} \max (\frac{l_{j - v}}{l_{j + v}}, \frac{l_{j + v}}{l_{j - v}})$ (2) where l_j is the length of the jth row of the foreground image, κ is the fixed value, which determines the image region used to calculate the shape segmentation operator, S (j, σ) is the difference of shape feature between the upper and lower σ rows with division line of the jth row.

Shape segmentation line h between different regions is obtained by shape segmentation operator. The final segmentation result f is obtained by the fusion of the color segmentation line and the shape segmentation line.

2.2 Similarity calculation

After obtaining the segmentation results and the region blocks, the weight of the edge is obtained to reflect the relationship between the blocks and the similarity between the pixels and the block. Two weighs ω^YY and ω^XY are defined. First, the relationship $ω_{ij}^{YY}$ between the region y_i and the region y_j is defined as $ω_{i j}^{Y Y} = {\begin{cases} \exp (\frac{c_{i} - c_{j}}{θ_{r}}), y_{i} \in N k_{x j} o r y_{i} \in N k_{x i} \\ 0, o t h e r w i s e \end{cases}$ (3) where Nk_x is the set of K-nearest neighbor of the region x, c_i is the average color information of the region R_i, $c_{i} = \frac{{\sum_{I_{k} \in}}_{R_{i}} I_{k}^{C}}{| R_{i} |}$ , $I_{k}^{C}$ is the color feature vector of the pixel k, |R_i| is the size of the region R_i, and θ_r is the weight coefficient.

Then the relationship ω^XY between the region Y_i and the pixel i is defined as $ω_{i j}^{Y Y} = P_{h i s t,} Y j (l = j | I_{i})$ (4) where P_{hist,Y_j} (l = j|I_i) is the posteriori probability calculated by constructing the histogram of the region Y_i.

The weight between pixels is calculated according to superpixel. The SIFT feature can describe the local features of the image, has invariance of scale change and rotation, and good stability for noise, angle change and affine transformation [12]. Because the image of busy waters in the actual scene are divided into several superpixels, and the size of the region is limited. In order to ensure that each region has SIFT feature points, the key point detection is abandoned and the dense sampling SIFT feature is used. Each pixel in the foreground is described in the unified direction and the unified scale, and is combined with the word package frame to represent the features of the superpixels [13]. First the DSIFT feature is extracted from the training data set. A dictionary needed to create the feature of a word packet is created by using the K-means clustering method. The SIFT feature descriptors corresponding to each pixel are mapped to the dictionary and the words corresponding to the descriptor are obtained. The frequency of each word in a super pixel is counted. Multiple superpixel histogram features constitute a superpixel feature description of a specific ship target.

The similarity of color features between targets is measured by using Bhattacharyya distance. Then the color distance between the sample A and the sample B is given by $d_{color} (A, B) = \sqrt{1 - \sum_{i} \frac{I_{A} (i) \cdot I_{B} (i)}{\sum_{i} I_{A} (i) \cdot I_{B} (i)}}$ (5) where I_A (i) and I_B (i) represents the color feature vector of the sample A and the sample B. The smaller the distance, the higher the similarity.

The same ship target obtained at different time and different location often has a change of pose, resulting in the drift of the feature of the superpixel histogram. If a simple Bhattacharyya distance is used and only the comparison of the superpixel histogram of the two targets, the mismatch rate will be increased. It can be avoided by using EMD crossover distance. The EMD distance is used to measure the similarity of the set and achieve the many-many matching. A set of superpixel features I of a specific ship target is expressed as a set of multi-feature set, that is, I = ((β₁, ω_β1), (β₂, ω_β2), ⋯, (β_n, ω_βn)), where β_i is the superpixel histogram vector, ω_βi is the weight of the vector β_i. Then the EMD distance between the superpixel features I_A = ((a₁, ω_a1), (a₂, ω_a2), ⋯, (a_n, ω_an)) of the target A and he superpixel features I_B = ((b₁, ω_b1), (b₂, ω_b2), ⋯, (b_n, ω_bn)) of the target B is defined as $d_{E M D} (A, B) = \min_{f_{i j}} \frac{\sum_{i = j}^{m} \sum_{j = 1}^{m} d_{i j} f_{i j}}{\sum_{i = j}^{m} \sum_{j = 1}^{m} f_{i j}}$ (6) where is the Euclidean distance of the vector a_i and the vector b_i, f_ij is the transport flow and satisfies the following constraints: ${\begin{array}{l} f_{i j} \geq 0 \\ \sum_{j = 1}^{n} f_{i j} \leq ω_{a i} \\ \sum_{j = 1}^{n} f_{i j} \leq ω_{b i} \\ \sum_{i = 1}^{m} \sum_{j = 1}^{n} f_{i j} = \min (\sum_{j = 1}^{n} ω_{a i} \sum_{j = 1}^{n} ω_{a i}) \end{array}$ (7)

The matching results of color feature and superpixel feature are fused to measure the similarity between objects. The distance of the target A and the target B is given by $d (A, B) = δ d_{color} (A, B) + λ d_{EMD} (A, B)$ (8) where δ and λ are the weights of the color feature and superpixel feature. The result given by Equation (8) is taken as ω^XY.

2.3 Energy function design

When the multi-layer graph structure is used for intelligent re-recognition of the ship target, the labeled sample ${\bar{D}}^{*}$ is first constructed according to the selected initial region. The recognition accuracy is improved by fusion of useful information from different region blocks in the partition library. The core idea of the algorithm is to estimate the pixel likelihood energy function of pixels and regions relative to the label l∈ { 0, 1 } by labeling the sample points to obtain the energy loss of a graph model [14]. The graph structure obtained by minimizing the energy loss is regarded as the recognition result.

According to the relationship between all pixels and the relationship between pixels and corresponding regions, the energy function related to pixel likelihood D_il and the label l is designed. $E_{l}^{X} = E_{l}^{PX} + E_{l}^{UX} + E_{l}^{HX}$ (9) where $E_{l}^{PX} = \sum_{i, j = 1}^{N} ω_{ij}^{XX} {(D_{il} - D_{jl})}^{2}$ is the label consistency constraint between adjacent pixels, used for 4 or 8 neighborhood modeling, $E_{l}^{UX} = \sum_{i = 1}^{N} φ_{i}^{X} d_{i}^{X} {(D_{il} - D_{il}^{*})}^{2}$ is the one-order constraint, used for label consistency of the labeled pixel in the learning process, $d_{i}^{X}$ is to adjust the relationship of the one-variable term and other two terms. $E_{l}^{HX} = \sum_{i = 1}^{N} μ d_{i}^{X} {(D_{il} - \bar{D_{il}})}^{2}$ is the consistency term of high order region, in this term, pixel likelihood and region likelihood must be similar. Regions often contain different labeled pixels and the parameter μ is to ensure the consistency of the region. $d_{i}^{X}$ is to adjust the relationship of the high order term and other two terms and $d_{i}^{X} = \sum_{j = 1}^{N} ω_{ij}^{XX}$ . The diagonal matrix $D^{X} = [d_{1}^{X}, \dots, d_{N}^{X}]$ and the weight matrix $W^{X} = {[ω_{ij}^{XX}]}_{N \times N}$ are defined. The parameter $φ_{i}^{X}$ is related with the seed. If i ∈ X^*, $φ_{i}^{X}$ isa fixed constant. Otherwise, $φ_{i}^{X} = 0$ . If i ∈ X^*, the seed likelihood $D_{ij}^{*} = 1$ . Otherwise, $D_{ij}^{*} = 0$ . Pixel likelihood ${\bar{D}}_{il}$ of the point i is estimated through region likelihood Z_l. The weight average of the corresponding region likelihood is defined. ${\bar{D}}_{il} = \sum_{k = 1}^{N_{Y}} P_{ik}^{XY} Z_{kl}$ , where $P_{ik}^{XY} = W_{ik}^{XY} / \sum_{j = 1}^{N_{Y}} W_{ij}^{XY}$ . In the proposed algorithm, a high order energy function term $E_{l}^{Y}$ for modeling the region likelihood model is defined in order to use the obtained over-segmented regions.

The region likelihood model Z_l is defined based on pixel likelihood. The energy function is defined as $E_{l}^{Y} = E_{l}^{PY} + E_{l}^{UY} + E_{l}^{HY}$ (10) where $E_{l}^{PY} = \sum_{i, j = 1}^{N_{Y}} ω_{ij}^{YY} {(Z_{il} - Z_{jl})}^{2}$ is the label consistency constraint between adjacent regions. If the color information of adjacent regions is similar, they should have similar labels. $E_{l}^{U Y} = \sum_{i = 1}^{N_{Y}} φ_{}^{i} d_{i}^{Y} {(Z_{i l} - Z_{i l}^{*})}^{2}$ is the two-variable constraint term. The label of each region is consistent with the interior pixel seed labeled information. $E_{l}^{HY} = \sum_{i = 1}^{N_{Y}} μ_{d} d_{i}^{Y} {(D_{il} - \bar{D_{il}})}^{2}$ is to ensure the average similarity of region likelihood and interior pixel likelihood, and $d_{i}^{Y} = \sum_{j = 1}^{N_{Y}} ω_{ij}^{YY}$ . The diagonal matrix $V^{Y} = [v_{1}^{Y}, \dots, v_{N_{Y}}^{Y}]$ and the weight matrix $W^{Y} = {[ω_{ij}^{YY}]}_{N_{Y} \times N_{Y}}$ are defined. The parameter $φ_{i}^{Y}$ is related with the seed. Y^* is the seed set of the region. For the region y_i, if ∀x ∈ y_i ∩ x ∈ X^*, y_i ∈ Y^*. Define the regional seed labeled vector ${\bar{Z}}^{*} = {[Z_{il}^{*}]}_{N_{Y} \times 1}$ . If y_i ∈ Y^*, the seed likelihood $Z_{il}^{*} = 1$ . Otherwise, $Z_{il}^{*} = 0$ . If y_i ∈ Y^*, $φ_{i}^{Y}$ is a fixed constant. Otherwise, its value is 0. Pixel likelihood D_xl, x ∈ y_i of the region is estimated through the likelihood $\bar{Z_{il}}$ of the region y_i. Define $\bar{Z_{il}} = \sum_{k = 1}^{N} P_{ki}^{XY} D_{kl}$ , where $P_{ki}^{XY} = W_{ki}^{XY} / \sum_{j = 1}^{N_{Y}} W_{ij}^{XY}$ and $P_{ki}^{XY} = P_{ki}^{YX}$ .

2.4 Intelligent re-recognition of specific ship target

Two energy functions $E_{l}^{X}$ and $E_{l}^{Y}$ have complementary characteristics. And the two energy functions are convex functions and optimized simultaneously. The graph structure obtained by minimizing the energy loss is considered as the re-recognition result.

The initial values are defined as $y_{l}^{*} = [\bar{D_{l}^{*}}, \bar{Z_{l}^{*}}]$ , where $\bar{D_{l}^{*}} = {[\bar{D_{il}^{*}}]}_{N \times 1}$ and $\bar{Z_{l}^{*}} = {[Z_{il}^{*}]}_{N_{Y} \times 1}$ , and $y_{l} = [\bar{D_{l}}, \bar{Z_{l}}]$ , where $\bar{D_{l}} = {[D_{ij}]}_{N \times 1}$ and $\bar{Z_{l}} = {[Z_{ij}]}_{N_{Y} \times 1}$ . The optimal solution of the energy loss of the graph structure is given by $(I - (I - Ψ)) \prod y_{l} = Ψ y_{l}^{*}$ (11) where the calculation matrix Ψ is $Ψ = [\begin{matrix} μ P^{X} (1 - \bar{μ}) P^{XY} \\ (1 - \bar{μ_{d}}) P^{YX} \bar{μ_{d}} P^{Y} \end{matrix}]$ (12) where P^X = (V^X) ^-1W^X, P^Y = (V^Y) ^-1W^Y, $P^{XY} = {[P_{ik}^{XY}]}_{N \times N_{Y}}$ , $P^{YX} = {[P_{ki}^{YX}]}_{N_{Y} \times N}$ , $\bar{μ} = \frac{1}{μ + 1}$ , and $\bar{μ_{d}} = \frac{1}{μ_{d} + 1}$ .

The current graph structure is the re-recognition result.

2.5 Incremental maintenance

For the infinite possible data, the problem of incremental processing is needed to be solved. When the ship image adds or removes an object in the corresponding target class of the data source, the corresponding class of the ship image in the data warehouse should add or delete an object [15 –17]. When the value of the corresponding object in the target class or the direct composite class of the data source is modified, the corresponding object in the corresponding class of the ship image in the data warehouse is needed to be modified. When the value of an object in the indirect composite class of the data source is modified, the value of the indirect composite class in the corresponding class in the data warehouse is needed to be modified.

For the object data type of object oriented database, the operation of adding or deleting in the object class or the indirect composite class will reflect as value modification of the composite parent object, so it cannot be executed. The incremental maintenance algorithm is as follows.

a. If the event create-object (MVTC, mobj) occurs, where MV is the ship image, mobj is the label of MVTC, then

The definition of all entity ship images and abstract ship images inherited by MV are obtained in a data warehouse and sorted in a partial order, V[1], V[2],…, V[n], where V[n] is MV.

for i: = 1 to n, execute:

The data types of all attributes of the V[i]’s attribute set are checked and divided into

The attribute set of the BT type BA[i],

The attribute set of the CT type CA [i],

The attribute set of the VT type VA [i].

Call function add_vobject (cid,V,BA,CA,VA,mobj).

Go to the Steph.

b. If the event delete_object (MVTC, mobj) occurs, the label mobj is deleted in the data warehouse and go to the Steph.

c. If the event update_object (MVTC, mobj) occurs, call update_object (mobj) and go to the Steph.

d. If the event update_object (MVTC, mobj) occurs and mcobj is an object of MVDC, then

The path expression path corresponding to MVDC is obtained from the data warehouse, and the reverse path expression inversepath is constructed according to the path.

The query requirement is constructed based on MVDC, mcobj, and inversepath to obtain the object set mobj_set in MVTC with the reference of mcobj.

The query is sent to the data integrator through the communication mechanism and the query result is accepted from the data integrator.

For each object mobj in mobj_set, call the function update_vobject (mobj).

Go to the Steph.

e. If the event update_object (AVTC, mobj) occurs and aobj is in the data warehouse, AV is the composite abstract ship image, aobj is the object of AVTC, then call the function update_vobject (aobj) and go to the Steph.

f. If the event update_object (AVTC, mobj) occurs and acobj is the object of AVTC, then

The path expression path corresponding to AVDC is obtained from the definition of composite abstract ship image the data warehouse, and the reverse path expression inversepath is constructed according to the path.

The query requirement is constructed based on AVDC, mcobj, and inversepath to obtain the object set aobj_set in AVTC with the reference of acobj and data source of AVC.

The query is sent to the data integrator through the communication mechanism and the query result is accepted from the data integrator.

For each object aobj in aobj_set, if aobj is in the data warehouse, call the function update_vobject (aobj).

Go to the Steph.

g. If the event update_object (VCC, mobj) occurs and cobj is in the data warehouse, cobj is the object of VCC, then

The query requirement is constructed based on VCC and cobj to obtain all the content of cobj.

The query is sent to the data integrator through the communication mechanism and the query result is accepted from the data integrator.

The value of cobj is modified in a data warehouse according to the query result. Note that if we want to modify the attribute values of the object data type, and the composite subobject of cobj exists, it is directly modified. Otherwise we need to load the composite subobject into the data warehouse (This process may be recursion).

h. Algorithm ends.

In the proposed algorithm, two functions add_vobject() and update_vobject () are called. The function add_vobject () is used to add an object to the object view. The function update_vobject () is used to modify an object in a view.

Considering the efficiency factor, the composite subobject will not be deleted when the algorithm deletes the object or modifies the composite subobject. This does not affect the correctness of the incremental maintenance algorithm or the front-end query results of the data warehouse. Regular running garbage-collecting program can delete those useless composite subobjects.

3 Results

3.1 Data set and experimental environment

In the experiment, the ship images in the Fleetmoom ship image library of busy waters in actual scene are taken for re-recognition. For the accuracy and reliability of the experiment, ship images were manually extracted and labeled as training and test samples, including 500 training samples, 310 test samples, and the resolution of each kind of the image is 200×200 pixels. Some image samples are shown in Fig. 2.

Fig. 2

Some image samples for experiment.

The experiment is carried out in the environment of MatlabR2014a, with Windows7 operating system, Intel Core i5, 4300U CPU with the main frequency 250 GHU, and memory 8 G. The SVM toolbox is the LIBSVM developed by Professor Lin Zhiren. The RBF kernel function is selected, in which the gamma parameter and the penalty coefficient are optimized by the grid.

3.2 Re-recognition test results

In order to verify the effectiveness of the proposed algorithm, the comparison between the proposed algorithm, the literature [6] algorithm, and the literature [8] algorithm is carried out. Figure 3 shows the results of the proposed algorithm. Figure 3(b-d) shows the similarity with the target. From Fig. 3, it can be seen that, for the specific ship target, after the monitoring scene and position change, the re-recognized target obtained with the proposed algorithm and the original target have high similarity. For non-original ship target, the similarity is very low. It shows that the proposed algorithm can effectively achieve the re-recognition of specific ship target.

Fig. 3

The recognition effect of the proposed algorithm.

Figure 4 shows the test results of the literature [6] algorithm. Figure 4(b-d) shows the similarity with the target. From Fig. 4, it can be seen that, for the specific ship target, for the case of the change of the monitoring scene and position, the similarity of the re-recognized target obtained with the literature [6] algorithm and the original target is basically the same as the non-original specific ship target. It shows that the literature [6] algorithm cannot realize the re-recognition of specific ship target.

Fig. 4

The recognition effect of the literature [6] algorithm.

Figure 5 shows the test results of the literature [8] algorithm. Figure 5(b-d) shows the similarity with the target. From Fig. 5, it can be seen that, for the specific ship target, under the condition of the change of the monitoring scene and position, the similarity of the re-recognized target obtained with the literature [8] algorithm and the original target is higher than the non-original specific ship target. It cannot realize the re-recognition of specific ship target.

Fig. 5

The recognition effect of the literature [8] algorithm.

3.3 Quantitative test of recognition accuracy

Quantitative test for Fleetmoom ship image database is carried out. The test index is selected as correlation coefficient, mean absolute error and average gradient.

Correlation coefficient:

Assume A (i, j) and B (i, j) are the gray values of two images, $\bar{A}$ and $\bar{B}$ are the means. The correlation coefficient is expressed as $C (A, B) = \frac{\sum_{i, j} [(A (i, j) - \bar{A}) \times (B (i, j) - \bar{B})]}{\sqrt{\sum_{i, j} {[(A (i, j) - \bar{A}) \times (B (i, j) - \bar{B})]}^{2}}}$ (13)

The correlation coefficient can reflect the correlation degree of two images. The correlation coefficient is closer to 1, the higher the proximity of the image, the higher the re-recognition precision.

Mean absolute error (MAE) is given by $MAE = \frac{\sum_{u \in U} \sum_{i \in {teseset}_{u}} | rec (u, i) - r_{u, i} |}{\sum_{u \in U} | {teseset}_{u} |}$ (14) where rec (u, i) is the recognition result, r_u,i is the actual result, teseset_u is the number of the test set. The lower the MAE, the higher the recognition the precision of the algorithm.

The average gradient reflects the contrast of minute details and texture changes in the image, which is defined as $Δ \bar{G} = \frac{1}{M \times N} \sum_{i = 0}^{M - 1} \sum_{j = 0}^{N - 1} \sqrt{(Δ I_{x}^{2} + Δ I_{y}^{2}) / 2}$ (15) where ΔI_x and ΔI_y are the difference in the direction of x and y. The larger the average gradient of the image, the smaller the contrast of the image details.

Quantitative test of re-recognition of Fleetmoom ship image database obtained with the proposed algorithm, the literature [6] algorithm, and the literature [8] algorithm is carried out. The results are shown in Table 1.

Table 1

Quantitative test results of the three algorithms

Index	The proposed algorithm	The literature [6] algorithm	The literature [8] algorithm
MAE	3.16	9.25	7.21
Correlation coefficient	0.95	0.76	0.72
Average gradient	92.35	75.15	81.33

From Table 1, it can be seen that, the MAE of the proposed algorithm is obviously lower than the literature [6] algorithm and the literature [8] algorithm, and the correlation coefficient and the average gradient are the highest compared with the other two algorithms. Therefore, the proposed algorithm has high re-recognition accuracy.

4 Discussion

In this paper, an intelligent re-recognition algorithm of specific ship target is proposed for the busy waters in the actual scene. Contributions of the proposed algorithm are as follows.

For a given ship to be recognized, first, whether it is in a data set is needed to be determined. If yes, it is needed to recognize the identity label in the data set. If no, it is needed to give a new identity label and add it to the data set. This will bring a problem. For the infinite possible data, the problem of incremental processing is solved by incremental maintenance.

For the newly observed ship to be recognized, how to adaptively judge whether it is in the data set is the technical difficulty for re-recognition in the busy waters under the real scene. In this paper, a graph structure is built and the recognition result is obtained by minimizing the energy loss, which is the recognition result, and the intelligence is high.

Key technologies of ship re-recognition include viewpoint invariant feature extraction and similarity measurement learning. The purpose is to solve the difference of apparent expression caused by different visual viewpoint changes under different cameras. In the re-recognition of the actual scene, the dense sampling SIFT feature is used to describe each pixel in the foreground according to the unified direction and the unified scale. Combined with the word packet framework, the SIFT description is used to represent the superpixel feature. On this basis, the similarity calculation is realized.

5 Conclusions

For the disadvantages of traditional algorithms, a new intelligent recognition algorithm of specific ship target in busy waters under the actual scene is proposed in this paper. Combining the existing feature extraction model and graph model, the graph structure is used to describe the identity relationship between the samples. Then segmentation of the target area is carried out. The energy loss of the graph model is obtained by designing the appropriate energy function. The graph structure obtained by minimizing the energy loss is regarded as the intelligent re-recognition result of the ship target. The recognition efficiency is increased by using incremental growth algorithm. Experimental results show that the proposed algorithm has high recognition precision.

The contents of further research are as follows.

Extension of the field of application. The proposed algorithm is not designed for specific task. Next, the proposed model will be applied to the fields of image recognition, target recognition, video analysis, common saliency, and target segmentation.

Modularization of algorithm implementation. Most of the researched contents are implemented with C++ or MATLAB and applied in the practical application. In order to promote the algorithm, we need to further modularize the proposed algorithm.

References

Kim

J.S.

and Jeong

J.S.

, Pattern recognition of ship navigational data using support vector machine, Microelectronics Reliability15(4) (2015), 268–276.

Arora

, Beilstein

M.A.

and Shippen

D.E.

, Evolution of arabidopsis protection of telomeres 1 alters nucleic acid recognition and telomerase regulation, Nucleic Acids Research44(20) (2016), 9821–9830.

Lihoreau

and Rivault

, Kin recognition via cuticular hydrocarbons shapes cockroach social life, Behavioral Ecology20(1) (2015), 46–53.

, Wang

, Xu

, Sun

, Huang

, Shen

., Stereospecific recognition and quantitative structure-activity relationship between antibodies and enantiomers, ofloxacin as a model hapten, Analyst140(4) (2015), 1037–1045.

Wang

, Gan

and Ma

, Under the nonlinear interference images of remote target automatic detection of the simulation of the ship, Computer Simulation32(5) (2015), 395–398.

Collins

, Ship detection performance using simulated dual-polarization radarsat constellation mission data, International Journal of Remote Sensing36(6) (2015), 1705–1727.

Yang

and Zhang

, A novel ship detection method for sar images based on nonlinear diffusion filtering and gaussian curvature, Remote Sensing Letters7(3) (2016), 210–218.

Berdnikova

, Klauson

, Mustonen

and Tommingas

, Underwater ship noise pattern detection and identification, Journal of the Acoustical Society of America142(4) (2017), 2686–2686.

Mroch

A.R.

, Laudenschlager

and Flanagan

J.D.

, Detection of a novel fh whole gene deletion in the propositus leading to subsequent prenatal diagnosis in a sibship with fumarase deficiency, American Journal of Medical Genetics Part A158A(1) (2015), 155–158.

10.

Wang

, Zhang

D.Q.

, Su

, Dong

J.W.

and Tan

S.K.

, Assessing hydrological effects and performance of low impact development practices based on future scenarios modeling, Journal of Cleaner Production179 (2018), 12–23.

11.

Xie

, Ge

, Jiang

, Liu

, Chen

, Wang

, Chen

, Qin

and Peng

, Study on biomolecules in extractives of Camellia oleifera fruit shell by GC-MS, Saudi Journal of Biological Sciences25(2) (2018), 234–236.

12.

Aladadi

W.M.

, Moustafa

M.F.

and Alruman

S.A.

, Genetic variability among seven cultivars of date palm (Phoenix dactylifera L.) based on embryonic DNA of old fruit, Kuwait Journal of Science45(1) (2018), 108–114.

13.

Pongnu

and Pochai

, Numerical simulation of groundwater measurement using alternating direction methods, Journal of Interdisciplinary Mathematics20(2) (2017), 513–541.

14.

Guang-Hui

, Yi

and Bing

, Raw material inventory optimization for MTO enterprises under price fluctuations, Journal of Discrete Mathematical Sciences & Cryptography20(1) (2017), 255–270.

15.

Freire

and Muatjetjeja

, Symmetry analysis of a lane-emden-klein-gordon-fock system with central symmetry, Discrete and Continuous Dynamical Systems-Series S11(4SI) (2018), 667–673.

16.

Awadalla

N.S.

, Hanna

M.A.

, Ismail

, Hassan

I.A.

and Elkhamisy

M.A.

, Period variation study and light curve analysis of the eclipsing binary gsc 02013- 00288, Mathematics and Nonlinear Sciences1 (2016), 321–334.

17.

Crespo

, Díaz–Toca

, Martínez

S.F.

and Lara

, Poisson and symplectic reductions of 4-DOF isotropic oscillators, The van der Waals system as benchmark, Mathematics and Nonlinear Sciences1 (2016), 473–492.