Side-scan sonar underwater target image sample amplification method based on CLL-sinGAN

Abstract

The invention discloses a CML-sinGAN-based side-scan sonar underwater target image sample amplification method, and belongs to the technical field of side-scan sonar image target recognition and amplification and deep learning. The method specifically comprises the steps that firstly, an eight-layer-scale pyramid network structure is designed according to the characteristics of a side-scan sonar image, and the diversity of a generated image is improved while image textures are fully learned; secondly, fusion is carried out in a generator, and information dispersion is reduced while target learning is enhanced; then, a loss function based on L1loss is introduced into the discriminator, the discrimination capability of the discriminator on the authenticity of the target is enhanced, and unreal false pictures are removed; and finally, based on the conversion model, carrying out high-quality amplification on the existing small sample side-scan sonar image, and carrying out YOLOv7 identification detection by using the amplified image. The method provides a new way for side-scan sonar small target sample amplification and a high-performance underwater target detection model.

Description

Technical Field

[0001] The present invention belongs to the technical field of side-scan sonar image target recognition and amplification and deep learning, and relates to a side-scan sonar underwater target image sample amplification method based on an improved sinGAN network. The present invention is an improved calculation method in the field of deep learning image amplification, and is applied to side-scan sonar image target recognition and image amplification. Background Art

[0002] As the global ocean development strategy goes deeper into the ocean, the demand for seabed topography detection is increasing, especially in the identification and detection of seabed targets, which plays a very important role in the fields of navigation safety, marine surveys, maritime search and rescue, military missions, etc. At present, ocean mapping mainly uses single-beam and multi-beam bathymetric systems and side-scan sonar systems. Among them, side-scan sonar has obvious advantages in seabed target identification with its high-resolution seabed acoustic imaging capabilities. Seabed target identification mostly relies on manual detection and identification, which has problems such as low efficiency, long time consumption, and strong subjective dependence. Therefore, it is very important to study methods for automatic detection of underwater targets.

[0003] Some scholars have adopted machine learning technology, combined with manual features and classification techniques, to achieve automated underwater target detection. However, these methods are limited in dealing with complex seabed environments because side-scan sonar images usually have problems such as low resolution, insufficient features, high noise, and deformation. The progress of deep learning technology in the field of computer vision has greatly improved the performance of target detection and has therefore been widely used in the field of underwater intelligent detection. Although the model based on deep convolutional neural network (DCNN) has significant effects, it has high requirements for high-quality training data, and side-scan sonar image samples are usually insufficient in number and limited in representativeness. Therefore, there is an urgent need to conduct sample enhancement research on side-scan sonar images of small sample underwater targets.

[0004] With the development of sample augmentation technology in the field of optical imaging, data augmentation technology in the field of underwater acoustics has also emerged. At present, there are two main methods for side-scan sonar image augmentation at home and abroad. One is the image style transfer method represented by GAN adversarial neural network, and the other is the image generation method based on the diffusion model. For example, Ye Xiufen used the AdaIN network for style transfer and achieved good results in target detection; Yang Zhiwei used the improved DDIM model for data augmentation and successfully improved the accuracy of the model; Huang Chao used the metalstyle network to perform data augmentation from the geometric and physical levels and obtained rich side-scan sonar images. However, both of the above models need to be based on a large number of side-scan sonar data sets and preprocessed. Various conditions increase the workload and limit the generation threshold.

[0005] Therefore, there are very few GAN networks that meet the requirements of few-sample data enhancement. In 2019, a network called SinGAN was proposed in a collaborative study between the Israel Institute of Technology and Google. This model can grasp the distribution characteristics of patches inside the image through in-depth learning of a single natural image. Through this learning, SinGAN can generate a series of image samples that are both high-quality and diverse. Therefore, the present invention selects the sinGAN network as the main method for side-scan sonar data enhancement for research. However, since the data sets of the original network are mostly colorful artistic style pictures, lakes and birds and other natural landscape pictures, when the sinGAN network was used to amplify the black-and-white side-scan sonar underwater target images, it was found that the targets were unreal and illogical. Summary of the invention

[0006] Based on the above situation, in order to enhance the network's learning of targets, this paper proposes a method for amplifying single image samples of side-scan sonar underwater targets based on CBL-sinGAN, targeting the characteristics of the side-scan sonar black-and-white waterfall image. First, an eight-layer pyramid network structure is designed based on the characteristics of the side-scan sonar image to fully learn the image texture while improving the diversity of the generated image; secondly, the Convolutional Block Attention Module (CBAM Moudlel) is integrated into the generator to enhance the learning of the target while reducing information diffusion; then, the loss function based on L1 loss is introduced into the discriminator to enhance the discriminator's ability to distinguish the authenticity of the target, eliminate unrealistic false images, improve the quality of the generated images, and improve the stability of the training, avoiding the collapse of the training mode. Finally, based on the conversion model, the existing small sample side-scan sonar images are amplified with high quality, and the amplified images are used for YOLOv7 recognition and detection. Experiments have proved that the side-scan sonar underwater target single image sample amplification method based on CBL-sinGAN proposed in the present invention can generate a large number of high-quality side-scan sonar amplified samples according to the characteristics of small number of side-scan sonar image samples, single color and diverse target morphology, providing a new way to solve the side-scan sonar small target sample amplification and high-performance underwater target detection model.

[0007] The technical solution of the present invention:

[0008] A method for amplifying side-scan sonar underwater target image samples based on CBL-sinGAN, the steps are as follows:

[0009] Step 1: Design an eight-layer multi-scale architecture based on sinGAN Moudle

[0010] The eight-layer multi-scale architecture based on sinGAN Moudle includes a generator pyramid and a discriminator pyramid. The generator pyramid is represented as {G0, G1, G2, G3, G4, G5, G6, G7}, and the discriminator pyramid is represented as {D0, D1, D2, D3, D4, D5, D6, D7}.

[0011] Assume {X0,X1,X2,X3,X4,X5,X6,X7} are images of different scales of the real side scan sonar image X. The generator pyramid performs pyramid training on image X at 8 scales and generates fake images respectively: in is the downsampled sample of X, N = 0 ~ 7, the downsampling factor is r N , r N Greater than 1. Generator G N The training object is a patch of a single image, not the entire image sample. Each generator generates a patch with the corresponding image X through adversarial training with the discriminator. N The patch distribution of the generator G is the most consistent with the real image sample. N The goal is to generate images of different scales to deceive the discriminator D N , discriminator D N The goal is to distinguish the generated samples The patch and the real sample x N The specific generation and training process of the patch is as follows:

[0012] Image generation starts from the coarsest scale, that is, from the eighth layer at the bottom, and is trained up layer by layer until the finest scale. Spatial noise {Z0, Z1, Z2, Z3, Z4, Z5, Z6, Z7} is added to each scale. The generator and discriminator at the same level have the same receptive field, and as the generation process proceeds, the receptive field decreases layer by layer. At the coarsest scale, the image is a pure generation process, that is, G7 maps the spatial Gaussian white noise z7 to the image sample.

[0013] The receptive field of the eighth layer is generally set to 1 / 2 of the image height, so the generator pays more attention to the overall global structure of the image. As the scale goes deeper, the receptive field setting becomes smaller and smaller, and the generator pays more attention to the detailed texture of the image. At the same time, in addition to the spatial noise z N In addition, the upsampled samples of the previous scale generated image are also used as the input of the generator Right now:

[0014] After the generator is trained step by step, fake side-scan sonar images of various scales are generated. and real side scan sonar images x N At the same time, it is sent to the discriminator D of each layer N The discriminator performs discrimination and feeds the result back to the generator. After adjusting the loss, the generator is retrained. The training ends when the discriminator cannot distinguish the authenticity of the image generated by the generator.

[0015] Step 2: Incorporate a CBAM-based generator into each pyramid

[0016] The key to the generator to generate high-quality images is to fully learn the target detail features and background features of the side scan sonar image. In order to enhance the learning of global information and local features in the input image and strengthen the interaction between channels and spatial dimensions, the present invention introduces the CBAM model (Convolutional Block Attention Module, CBAM Moudle) in the second layer of the generator and places it behind the body layer of the generator, such as Figure 2 shown.

[0017] Figure 2 The overall architecture after adding the attention mechanism CBAM module is given. It can be seen that the CBAM model consists of two independent sub-modules: the channel attention module CAM (Channel Attention Module) and the spatial attention module SAM (Spatial Attention Module). Compared with the attention mechanism that only focuses on space, better results can be achieved. This can save parameters and computing power while being easily inserted into various levels of the network. This module aims to reduce the diffusion of information while amplifying the cross-dimensional interactions of the image channels and space, thereby improving network performance. By focusing on relevant features and minimizing interference, the CBAM module makes the side-scan sonar image more delicate and detailed, which is crucial for producing realistic and high-quality output images.

[0018] Step 3: Add a loss function based on L1loss to the discriminator module

[0019] The process of training the generator is carried out from the coarsest scale to the finest scale, training the multi-scale architecture layer by layer, and keeping the generator fixed once each layer is trained. The training loss of the generator of the Nth layer includes an adversarial term and a reconstruction term: Among them, L adv (G N ,D N ) represents the adversarial loss, Lrec (G N ) represents the reconstruction loss, and α represents the hyperparameter.

[0020] (1) Adversarial term:

[0021] Each generator G N Both are associated with a Markov discriminator D N Pairing, D N The overlapping patches of the input are discriminated. Based on the original loss function WGAN-GP, we introduce L1 loss for discrimination, which further increases the stability of training, and the final discrimination score is averaged. The loss of the entire image is defined here instead of the loss of a single patch, so that the network can learn the boundary conditions.

[0022] The formula for adversarial loss is: Ladv(GN,DN)=L WGAN-GP +L1 loss

[0023] Where L WGAN-GP is the loss function of the original network, L1 loss is the added l1 loss function.

[0024] L1 loss is a mean square error loss that focuses on the target, because these false targets will have a greater contribution after the square loss, and thus feedback to the generator, "promoting" the generator to generate high-quality, highly representative samples that are more consistent with the side-scan sonar target image. This paper adds L1 loss to the discriminator, and calculates L1 loss while calculating WGAN-GP, to enhance the discriminator's ability to distinguish false targets. L1 loss The calculation is as follows:

[0025] Where f(x i ) and y i They represent the predicted value of the i-th side-scan sonar false image generated at the same scale and the corresponding real side-scan sonar image, and n is the number of generated images.

[0026] (2) Reconstruction term:

[0027] To ensure that after inputting a specific set of noise maps, the generator can generate the original image based on this noise map. Let the image generated and reconstructed after the Nth generation be The random Gaussian white noise generated by this layer is Z N , then for N<7, we have:

[0028] For N=7, the reconstruction loss is calculated as: L rec =||G7(z7)-x7|| 2 .

[0029] Among them, the spatial noise Z of each layer N The standard deviation σ N Reconstructed image and the true image x N The decision can indicate the number of details of the image that need to be added for training at this level. The specific calculation method is as follows:

[0030]

[0031] Step 4: Use the CBL-sinGAN model constructed in steps 1 to 3 to amplify the side scan sonar underwater target image samples.

[0032] Beneficial effects of the present invention: The advantages of CBL-sinGAN are as follows:

[0033] (1) It can retain the style, background texture, and integrity of the target image of the original side-scan sonar image to the greatest extent possible, and can generate and amplify more realistic side-scan sonar images.

[0034] (2) It can simulate the distortion of objects caused by factors such as seabed noise and water disturbance by performing operations such as deformation and blurring of objects, thereby generating images that are more consistent with the measurements obtained under the actual operating conditions of side-scan sonar.

[0035] (3) For datasets with too few samples (tens to hundreds), which are not enough to form a fixed style, amplification has a huge advantage. Traditional networks cannot form a fixed style, and the generated side-scan sonar images are not realistic.

[0036] (4) The relationship between the target and the shadow can be taken into account when generating images. This is a unique feature of side-scan sonar images. The original network does not perform well in shadow generation when generating images. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the eight-layer scale pyramid structure of sinGAN;

[0038] Figure 2 It is a generator network structure improved based on CBAM;

[0039] Figure 3 It is a partially realistic side-scan sonar dataset;

[0040] Figure 4The style transfer effect of the cycleGAN network on a small data set, where (a) shows the effect of CycleGAN on optical image style transfer, and (b) shows the effect of three different networks on 3Dmax modeling images.

[0041] Figure 5 It is the effect of pyramid experiments at different scales;

[0042] Figure 6 It is the effect of converting some side scan sonar images using different modules. DETAILED DESCRIPTION

[0043] The experiment of the present invention is described in detail below in conjunction with the accompanying drawings:

[0044] like Figure 1 As shown in Figure 2, our model consists of a GAN pyramid, where both the training process and the generation process are completed in a coarse-to-fine manner. At each scale, the generator G N It learns and is responsible for generating images of different scales in a coarse-to-fine manner. Its input is noise Z and upsampled samples of the previous level output. N The function of is to determine whether its input sample is a real sample or a forged sample.

[0045] It should be noted that the generator G N The training objects are patches of a single image, not the entire image sample. In order to handle side-scan sonar targets, such as shipwrecks, aircraft, etc., the generator needs to capture global properties such as the layout and shape of large objects in the image, such as the hull structure of a shipwreck, the wing features of an aircraft, etc. To achieve this, the generator framework consists of a series of hierarchical patch-GANs (Markov discriminators), each responsible for capturing the distribution of patches at different scales of the training image. These patch-GANs have small receptive fields and limited capacity to prevent memorizing a single image. Although traditional GANs have explored similar architectures, and sinGAN has been applied in many fields for image augmentation generation, we are the first to use it for side-scan sonar image augmentation of a single image.

[0046] In order to verify the performance of the CBL-sinGAN-based side-scan sonar underwater target single image sample amplification method in target detection, the present invention designed a variety of comparative tests, amplified shipwreck targets, selected 215 shipwrecks of various shapes and containing different background information, 62 airplanes, 115 fish schools, and 2 underwater frogmen, and amplified 2650 amplified images using the network. We conducted some experiments using these amplified images and other real side-scan sonar shipwreck images. Among them, the hardware information used for model training is Core TMi7-13700KF CPU and NVIDIA GeForce RTX 4070GPU 12GB. The software compilation environment used is pytorch1.6.0, cuda11.8, python3.10 under windos10, some data sets are as follows Figure 3 shown.

[0047] At present, the mainstream style transfer networks are based on a large number of style pictures to learn image texture and style, such as cycleGAN, WGAN, etc., which have achieved good results in the field of style transfer. Some scholars have also introduced it into the image amplification of side scan sonar and achieved good results. However, the training of this network is based on a large amount of data (more than 200 images). When the image data is small (dozens or about a hundred images), the transfer effect is not ideal. Figure 4 shown.

[0048] The CBL-sinGAN-based side-scan sonar underwater target single image sample augmentation method is essentially a process of rearranging and combining real side-scan sonar images, and the image is augmented by reorganizing the image background, stretching and scaling the target shape, transforming the number of targets, transforming the target position, etc. Since sinGAN is a new field for processing side-scan sonar images, we found many problems during the model training process and made corresponding improvements:

[0049] Selection of scale when generating images: Unlike artistic style images, side scan sonar images need to maintain the stability and authenticity of the target when amplifying to prevent the generated images from having unrealistic targets. Experiments have found that the choice of training scale has a great impact on the quality of image generation during training. At a lower scale, the network focuses on learning the details of the image, and does not learn the overall target sufficiently, resulting in the generated image target being fragmented, with only the basic texture of the image. As the training scale increases, the target learning gradually takes shape. However, the consequence of the increase in scale is an increase in training time, and if the training scale is set too much, the diversity of generated images will decrease. The calculation of relevant indicators is shown in Table 1.

[0050] Table 1 shows the corresponding parameters of the pyramid experiment at different scales

[0051]

[0052] According to the table analysis in Table 1, when scale = 8, the FID index of the generated image performs best, indicating that the model can learn the style of the side scan sonar image well at this time; at the same time, the IS index is 5.583, which is also excellent, indicating that the diversity of the generated image is also very good. The reason for the higher diversity when scale = 3 may be that it only learns the texture of the image, regardless of the overall image target, so there is greater diversity; in terms of MMD index performance, 0.24 is also in a relatively good range, and within the statistical range, the generated image and the original image are also very similar. Therefore, considering all factors, the number of structural layers of the CBL-sinGAN pyramid network is set to 8 layers. At this time, the model can fully learn the image texture while improving the diversity of the generated image.

[0053] Ablation experiment and evaluation: In order to verify the role of each module in the performance of the model of the present invention, the control variable method is used to perform ablation experiments on the CBAM module and L1 loss function respectively, and the evaluation indicators are FID, MMD and IS. Four groups of control experiments are designed with the same experimental configuration and training data set. The experimental results are shown in the table in Table 2.

[0054] Table 2 is the comparative analysis data of the ablation experiment

[0055]

[0056]

[0057] It is not difficult to observe from Table 2 that compared with the control group (Group 1), Group 2 and Group 4, after incorporating the CBAM module, help the model focus more on the key information in the image by focusing on important channels and spatial regions, thereby improving the expressiveness of features. This is reflected in the improvement of the indicators FID and IS. Comparing Group 1 and Group 3, it can be seen that the improvement of the IS indicator shows that L1 Loss performs well in processing data sets with outliers or performing feature selection. Combination of CBAM and L1 Loss: The results of Group 4 show that the combination of CBAM and L1 Loss not only improves the quality of the image (lower FID and higher IS), but also increases the diversity of the generated images (lower MMD). This is because CBAM improves the expressiveness of the features, while L1 Loss increases the robustness of the model to outliers. This enables the model to amplify the diversity of the image while ensuring the authenticity of the image.

[0058] The conversion effect of some side scan sonar images using different modules is as follows Figure 6As shown in the figure. It can be seen from the figure that the image generated by group 1 has the lowest authenticity and does not conform to the normal target structure. For example, the two upper bodies of people overlap, the virtual shadows of the ship overlap, the target of the aircraft is blurred, and the school of fish pays too much attention to the learning of the background and ignores the learning and expansion of the target, resulting in a large area of ​​black in the generated image. At the same time, it can be found that the shadow of the generated image does not correspond to the target, and there is only a shadow without a target. After adding CBAM to group 2, it can be seen from group 1 that the model has a certain improvement in the learning ability of the target when generating the target, but it can be seen from the target of the shipwreck that there is still a phenomenon of virtual shadow overlap. Comparing group 3 and group 1, it can be seen that the model has improved its ability to maintain the shape of the target when generating, which effectively prevents the distortion of the target, but the diversity of the model has not increased, and the shape of the underwater frogman has not changed. By comparing Group 4 and Group 1, it can be seen that after integrating the CBAM and L1 loss functions, the model has achieved very good results in learning the texture details of the target, maintaining the posture, and amplifying the background. It generates multiple samples with complete details and less distortion. At the same time, the relationship between the target and the shadow can also be well maintained in one-to-one correspondence.

[0059] Performance of the model in target detection: Considering that the purpose of the present invention is to amplify the side-scan sonar underwater target sample image to improve the performance of the target detection model based on deep learning, the present invention selects a target detection model based on deep learning for comparative experiments. There are many detection models at present. The YOLOv5 detection model is a lightweight, fast, and mature detection model that is suitable for this experiment, so the YOLOv5 detection model is selected for deployment.

[0060] The target image is a shipwreck, and three sets of data sets are designed to train and deploy the YOLOv5 model, including only real shipwreck images, only augmented shipwreck images, and both real and augmented shipwreck images. 100 real side-scan sonar shipwreck images are selected to evaluate the performance of the detection model. The details are shown in the table. The augmented shipwreck image data has been screened to remove images with poor generation quality.

[0061] Table 3 Composition of YOLOv5 detection model dataset and validation set

[0062]

[0063] The trained model was tested using 100 real side-scan sonar shipwreck target images. The evaluation indicators used were precision, recall, and average precision (AP), which are widely used in the field of target detection. The test results are as follows:

[0064] Table 4 Detection effect of different training sets on real measured side scan sonar sunken target images

[0065]

[0066]

[0067] It can be seen from the table that, by comparing YOLOv5-1 and YOLOv5-2, it can be seen that the model trained by amplifying the sunken ship target of the side-scan sonar using the method of the present invention is higher in precision, recall and average precision than the model trained only using the real side-scan sonar image, proving the key role played by the amplification map in improving the model performance. By comparing YOLOv5-2 using only the amplification map and YOLOv5-3 using the amplification map and the real side-scan sonar image, it can be seen that the differences in their various evaluation indicators are not large, proving that the improvement in model performance is mainly due to the side-scan sonar amplification data generated by the method of the present invention, further illustrating that the amplified image of the present invention can meet the authenticity and diversity requirements of the side-scan sonar image.

[0068] In order to prevent the deviation of the data set on a single detection model, multiple detection models (YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5x) were used to conduct precision comparison experiments. The experimental analysis results show that due to the complexity of the detection model and the difference in training time, there are differences in precision between different detection models. However, when comparing the three experimental groups horizontally, the precision rates of group 3> group 2> group 1. This further proves the effectiveness of the experimental data.

[0069] Table 5 Precision comparison between different models

[0070]

Claims

1. A method for amplifying underwater target image samples of side-scan sonar based on CBL-sinGAN, characterized in that: Here are the steps: Step 1: Design an eight-layer multi-scale architecture based on sinGAN Moudle The eight-layer multi-scale architecture based on sinGAN Moudle includes a generator pyramid and a discriminator pyramid, the generator pyramid is represented by {G0, G1, G2, G3, G4, G5, G6, G7}, and the discriminator pyramid is represented by {D0, D1, D2, D3, D4, D5, D6, D7}; Assume {X0,X1,X2,X3,X4,X5,X6,X7} are images of different scales of the real side scan sonar image X. The generator pyramid performs pyramid training on image X at 8 scales and generates fake images respectively: in is the downsampled sample of X, N = 0 ~ 7, the downsampling factor is r N , r N Greater than 1; Generator G N The training object is a patch of a single image, not the entire image sample; each generator generates a corresponding image X through adversarial training with the discriminator. N The patch distribution of the generator G is the most consistent with the real image sample. N The goal is to generate images of different scales to deceive the discriminator D N , discriminator D N The goal is to distinguish the generated samples The patch and the real sample x N The specific generation and training process of the patch is as follows: Image generation starts from the coarsest scale, that is, from the eighth layer at the bottom, and is trained upward layer by layer until the finest scale. Spatial noise {Z0, Z1, Z2, Z3, Z4, Z5, Z6, Z7} is added to each scale respectively; the generator and discriminator at the same level have the same receptive field, and as the generation process proceeds, the receptive field decreases layer by layer. At the coarsest scale, the image is a pure generation process, that is, G7 maps the spatial Gaussian white noise z7 to the image sample In addition to the spatial noise z N In addition, the upsampled samples of the previous scale generated image are also used as the input of the generator Right now: After the generator is trained step by step, fake side-scan sonar images of various scales are generated. and real side scan sonar images x N At the same time, it is sent to the discriminator D of each layer N After the discriminator is trained, the result is fed back to the generator, and the generator is retrained after the loss is adjusted. The training ends when the discriminator cannot distinguish the authenticity of the image generated by the generator. Step 2: Incorporate a CBAM-based generator into each pyramid Introduce the CBAM model in the second layer of the generator and place it behind the body layer of the generator; The CBAM model consists of two independent sub-modules: the channel attention module CAM and the spatial attention module SAM; Step 3: Add a loss function based on L1loss to the discriminator module The process of training the generator is carried out from the coarsest scale to the finest scale, training the multi-scale architecture layer by layer. Once the generator of each layer is trained, it remains fixed; the training loss of the generator of the Nth layer includes an adversarial loss and a reconstruction loss: Among them, L adv (G N ,D N ) represents the adversarial loss, L rec (G N ) represents the reconstruction loss, α represents the hyperparameter; (1) Fighting against losses: The formula for adversarial loss is: Ladv(GN,DN)=L WGAN-GP +L1 loss Where L WGAN-GP is the loss function of the original network, L1 loss is the added L1 loss function; L1 loss The calculation is as follows: Where f(x i ) and y i They represent the predicted value of the ith side-scan sonar false image and the corresponding real side-scan sonar image generated at the same scale, respectively, and n is the number of generated images; (2) Reconstruction loss: Suppose the image generated and reconstructed after the Nth generation is The random Gaussian white noise generated by this layer is Z N , then for N<7, we have: For N=7, the reconstruction loss is calculated as: L rec =||G7(z7)-x7|| 2 ; Among them, the spatial noise Z of each layer N The standard deviation σ N Reconstructed image and the true image x N The decision can indicate the number of details of the image that need to be added for training at this level; the specific calculation method is as follows: Step 4: Use the CBL-sinGAN model constructed in steps 1 to 3 to amplify the side scan sonar underwater target image samples.

Images