In order to train the model for deep learning, a lot of data is typically required. However, in practice, certain data sets are challenging to collect, leading to a small quantity of data in this category that falls short of the required data volume for deep learning. Experts have thus suggested data augmentation approaches to successfully address this issue (Najafabadi et al., 2015). The data augmentation techniques such as random rotation and mosaic was used. Given training data, mosaic randomly crops four images and stitch them together to create one. It has the advantage of enriching the background of the image and enhancing the batch size discretely so that it can help to minimize the model dependence on a large batch size when training ( Figure 1 ).

The images were annotated using the Labellmg software, which generates a json annotation file and transforms it into a txt file. The type and number of the labeling target, the labeling image’s standardized width and height, and the center point’s coordinates are all information that can be found in the txt file. Figure 2 displays the labelling outcomes.

YOLO is a regression-based technique and, despite being less accurate, is actually faster than region proposal-based methods like R-CNN (Girshick et al., 2014). The goal of YOLO is to achieve object identification by approaching it as a regression and classification issue. Identifying the bounding box coordinates for the objects in the images is the first step, and second step is to classifying the objects that are identified in a class. This is accomplished in a single step by first splitting the input images into a grid of cells, then determining the bounding box and relative confidence score for each cell’s containing object.

The YOLOv5 network is one of the recent research advancements in the YOLO series of algorithms. Despite sharing a network structure with the YOLOv4 network, it is smaller, has a faster running speed and convergence speed, and uses a lightweight algorithm. Additionally, it improves precision at the same time. As a result, the YOLOv5s algorithm is used in the current study work to detect ships in SAR images. Four components make up the YOLOv5 network structure: input, backbone, neck, and prediction. The Yolov5 framework architecture is displayed in Figure 3A . Networks can be categorized as YOLOv5l, YOLOv5m, YOLOv5x, and YOLOv5s. Their widths and depths may differ significantly, but their network structures are comparable. The network structure of YOLOv5s is the shortest, shallowest, runs the fastest, and has the least accuracy. As a result, the accuracy continues to rise, the speed of operation declines, and the other three network structures increasingly deepen and widen. Adaptive anchor box operation, mosaic data augmentation, image scaling, and CSP structure were used to process the input dataset, while focus and CSP framework were used to build the backbone. Focus increased network speed and cut down on floating-point operations (FLOPs) by clipping the input image. Figure 3A presented the focal structure. The two CSP (Wang et al., 2020a) framework that were used by YOLOv5 were CSP1_X and CSP2_X; CSP1_X was utilized for down sampling in the backbone while CSP2_ X was utilized in the neck. CSP can reduce operations while increasing the network’s capacity for learning and guaranteeing accuracy. Figure 3B depicted the two CSPs’ structures; the neck used the SPP-net and FPN + PAN framework to improve the network’s feature fusion effect, while the prediction employed the GIOU_ Loss (Rezatofighi et al., 2019), which did not only focus on the overlap between the prediction box and the ground truth but also on the non-overlapping areas. (Yu et al., 2016) found that GIOU maintains the benefits of IOU while solving its issues. The computation for Equation (1-2) is as follows:

Four separate networks were used, and YOLOv5 was continuously upgraded as well. Version 5.0 of YOLOv5s was used in this study project; in comparison to version 4.0, this version modified all functional activation in the framework to SiLU (Elfwing et al., 2018), eliminated the conv in the CSP, and designated it C3 as presented in Figure 4 . Additionally, v5.0 has a smaller and faster network structure than v4.0.

In this section of the study, the improvements made to the YOLOv5 classifier have been described in accordance with the guidelines of the proposed research challenge. The neck and backbone parts are enhanced to produce the greatest identification outcomes.

SSDD dataset: The first and most important stage in ship detection using deep learning techniques is the construction of a sizable and representative dataset. Therefore, the experiment utilize the SSDD (Li et al., 2017a) dataset, which have 1160 SAR images with resolutions ranging from 1m to 10m with polarizations in HH, HV, VV, and VH from RadarSat-2, TerraSAR-X, and Sentinel-1. Each sample image has a dimension of roughly 800 x 800, with a ratio of 7:1:2, where the SSDD dataset is divided into three sets for the experiment: a training set, a validation set, and a testing set ( Table 1 ).

AirSARship Dataset: In the present study, experiments also use the AirSARShip-1.0 (Xian et al., 2019) dataset to assess the performance of proposed model utilize high-resolution SAR ship identification dataset. Gaofen-3 provided 31 single-polarized SAR images for AirSARShip-1.0. Most images have a size of 3000 x 3000 pixels, while one has a size of 4140 x 4140 pixels with resolutions ranging from 1 to 3 meters with HH polarization. The large scene image has been split into 1000 x 1000 slices with a ratio of 7:1:2, where the dataset is divided into three sets for the experiment: a training set, a validation set, and a testing set ( Table 1 ).

Two SAR large scene images from the Chinese GF-3 satellite, as shown in Figure 13 , further illustrate the efficacy of the suggested strategy for identifying different size ships in SAR large scene images with complicated sceneries. These images contain inshore and offshore scenery as well as ship targets at various scales. In Figure 8 , some image slices are presented and offshore and inshore scenes as well as multiscale ship targets are primarily shown in Figure 8A . The dataset clearly shows that both off-shore and inshore scenarios are included, and that the sizes of the ships fluctuate widely.

The experiments are all carried out using PyTorch 1.7.0, CUDA 10.1, and CUDNN 7.6.5 on an NVIDIA Geforce GTX 2080Ti GPU and an Intel Core i9-9900KF CPU. The PC, which was equipped with a deep learning environment for our research and was running Windows 10, is depicted in Table 2 . Additionally, each model was trained over 100000 iterations utilizing the Stochastic Gradient Descent (SGD) technique on a total of two images per minibatch. The initial learning rate was set at 0.001, while the weight decay was set to 0.00004. In every trial, the detection threshold IOU was set to 0.7. Table 2 displayed the experimental hardware and software configuration. During the experiment, the same platform was used for all comparison techniques.

Since optical and SAR image object detection tasks are comparable, the effectiveness of various approaches is assessed using a variety of established indicators, such as average precision (AP), recall rate (r), precision rate (p), F score (F1), and these indications are specifically formulated in following equations (3-6):

The number of correctly recognized ships, false alarms, and missing ships are denoted by the acronyms FN (false negative), FP (false positives), and TP (true positives). The precision and recall are combined into the F1 score as follows:

The complete detection effectiveness of the various models is assessed using the AP and F1-score metrics, and a higher number indicates a superior detector performance.

The percentage of ground truth ships that networks correctly predict in all predictions is referred to as the precision rate. The percentage of ground truth ships that networks correctly predicted in all ground truth ships is referred to as recall rate. F1 is a comprehensive statistic that combines precision rate and recall rate to assess the effectiveness of various framework. AP outlines the region beneath Precision-Recall (PR) curves and also shows the overall effectiveness of various approaches. Additionally, Frames-Per-Second (FPS), which is derived from Equation (7), is used to assess the detection speed of various approaches. A method achieves a higher speed the higher the FPS.

When processing an image, the inference time (Tper-imgis) is the cost of a method.

In this section, the proposed approach and alternative CNN-based approaches, such as Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), R2CNN (Jiang et al., 2017), ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN (Lu et al., 2019), Cascade R-CNN (Cai and Vasconcelos, 2018), YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free-Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and yolov5-X, are assessed using offshore and inshore ships of testing sets. Aside from these metrics, F1, AP, and FPS are also employed to investigate the applicability of various methodologies. The suggested model’s identification performance against other CNN-based approaches tested on offshore ships and inshore ships based on SSDD dataset and AirSAR ship dataset is presented in Tables 3 , 4 ; Figures 9 , 10 . The current model provides the best accuracy for offshore SSDD (about 95.36% AP for the offshore scenes). The second-best result is 89.03% from the R2CNN approach, although it still performs better at detecting anomalies than the currently suggested model by 6.33% AP. The studied model also delivers the best accuracy for inshore on SSDD (about 92.27% AP for the inshore scenes). The second-best result is 83.53% from the R2CNN approach, although it still performs better in terms of detection than the currently suggested model by 8.74% AP. The researched model provides the best accuracy for offshore on AirSARship (about 94.57% AP for the offshore scenes). It has the best detection performance, as seen by the second-best result of 88.27% from the Quad-FPN approach, which is still 6.3% AP less than the proposed model. The proposed model also delivers the best accuracy for inshore on AirSARship (about 91.11% AP for the inshore scenes). It also has the best detection performance, with the second-best estimate coming from the Lite-Yolov5 approach at 84.94%, however it is still 6.17% AP lower than the current model.

The suggested model and the other state-of-the-art CNN based approaches include Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), R2CNN (Jiang et al., 2017), ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN (Lu et al., 2019), Cascade R-CNN (Cai and Vasconcelos, 2018), YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free-Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and yolov5-X techniques all have detection accuracies that are higher for offshore scenes than for inshore situations. This is reasonable considering that the former has a more complicated background than the latter. Perhaps as a result of their poor small ship identification capabilities, the other alternative approaches have lower precision values than the suggested model. In this current research work, the recall values of the proposed model are occasionally less than those of other offered methods. As a result, a suitable score threshold can be further thought about in the future to balance missed detections and false alarms. Additionally, the current model appears to be faster than other approaches based on the FPS data in Tables 3 , 4 , and Figures 9 , 10 , potentially as a result of the separable depth-wise and point-wise convolutions utilized in the backbone network. In conclusion, the offshore scene has greater accuracy, AP, and F1 scores for both datasets than the inshore scenario. This might be due to the inshore scene’s densely packed ships and increased backdrop interference from the land. Additionally, it demonstrates that it is more difficult to spot ships in the inshore scene than it is in the offshore environment.

The proposed model’s performance in terms of detection results compared to existing convolutional neural network -based approaches tested on offshore ships and inshore ships using SSDD dataset and AirSARship dataset is shown in Figures 11 , 12 . The suggested model is capable of detecting different SAR ships with multiscale sizes under varied backgrounds. This demonstrates its great scale and scene adaption together with excellent detection performance. The currently proposed model can increase the detection confidence scores when compared to the second-best CNN-based ship detector R2CNN. For instance, the suggested model raises the confidence score in Figure 11 first detection sample from 0.96 to 1.0. This can demonstrate the better trustworthiness of the newly proposed model. It is evident that the inshore scenario contains a sizable number of tightly packed ship targets. The other suggested solutions miss certain closely grouped inshore ships. The proposed model, however, is capable of accurately localizing and detecting these multiscale ships with high probabilities.

The other suggested CNN-based techniques in this research can precisely identify the port’s heavily docked ships. However, it can be observed that the suggested method is more accurate and can detect these ships better when comparing the detection outcomes of several proposed CNN-based systems. The detection outcomes of the Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), R2CNN (Jiang et al., 2017), ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN (Lu et al., 2019), Cascade R-CNN (Cai and Vasconcelos, 2018), YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free-Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and yolov5-X algorithms contain several false alarms for the offshore and inshore scene ( Figures 11 , 12 ). Additionally, there are a few missed ships in the detection findings, which could be a result of how closely docked the ship targets are, making it more challenging for the framework to discriminate between them. Similarly, it is observed by comparing the suggested model’s detection results that they are more precise than those produced by existing CNN- based techniques. Figures 11 , 12 displays the outcomes of various CNN-based object detection techniques in an offshore scenario created for SAR images. It is evident that the offshore landscape contains a substantial number of dense multi-scale ship targets (the first two column of Figures 11 , 12 ).

A pink color circle denotes false alarms in the identification outcomes of other proposed CNN-based models in this study that are not the current model. This might be due to a small number of false alarms that closely resemble ships, creating it great challenging for the network to successfully recognize. Because there are wakes of ships and surroundings, such ship A in Figure 12 , the Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), R2CNN (Jiang et al., 2017), ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN (Lu et al., 2019), Cascade R-CNN (Cai and Vasconcelos, 2018), YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free-Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and yolov5-X algorithms can distinguish between wakes of ships and their surroundings. Additionally, several ships are overlooked by the Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), R2CNN (Jiang et al., 2017), ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN(Lu et al., 2019), Cascade R-CNN (Cai and Vasconcelos, 2018), YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free-Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and yolov5-X algorithms, as seen by the red circles in Figures 11 , 12 . The inability to extract distinguishing characteristics of ships and interference may be to blame. However, the suggested model could identify these ships without any false alarms, which is also one of the driving forces for this paper and highlights the current proposed model’s powerful and robust feature representation skills.

The suggested model is compared with other state-of-the-art CNN based methods SAR ship identification techniques. As can be observed, two-stage or multi-stage approaches typically execute detection tasks more effectively than single-stage methods. However, compared to these two stage or multistage detection methods, one stage detection methods clearly have a faster inference efficiency. This might be as a result of the two-stage detection network’s sophisticated network architecture and increased computational load. Some one-stage detection techniques that are more effective at detecting ships have recently been proposed, including YOLOv4-LITE (Liu et al., 2022), EfficientDet (Tan et al., 2020), Free Anchor (Zhang et al., 2019), Lite-Yolov5 (Xu et al., 2022), and two-stage detectors like ARPN (Zhao et al., 2020), DAPN (Cui et al., 2019), Quad-FPN (Zhang et al., 2021b), HR-SDNet (Wei et al., 2020), Grid R-CNN(Lu et al., 2019), and Cascade R-CNN (Cai and Vasconcelos, 2018) are the performance comparisons of the current model with other cutting-edge detectors.

The detection accuracies of inshore scenes for the proposed model and the other CNN-based techniques are also all lower than those of offshore scenes. De-formable convolution can lessen the interference of complicated backgrounds, particularly for inshore sceneries, hence the recently presented technique seems to be robust to background interferences. The other state-of-the-art techniques are less precise and have lower recall values than the model now under study because of their poor small ship recognition capabilities. As a result, it will be possible to consider an acceptable score threshold in the future to balance missed detections and false alarms. Additionally, accuracy needs to be further improved, for example, when striking military targets with precision. It might be suggested in the future to choose between speed and accuracy. The suggested model has a higher detection effectiveness. This might be the case because other methods overlook smaller targets since they do not consider the underlying data in the prediction layer. There are some false alarms in the identification outcomes of previous techniques for the complicated inshore scenarios. Particularly, several land features in the inshore scene are wrongly identified as targets by Faster-RCNN, SSD, R2CNN, ARPN, DAPN, Quad-FPN, HR-SDNet, Grid R-CNN, Cascade R-CNN, YOLOv4-LITE, EfficientDet, Free-Anchor, Lite-Yolov5. By doing several experiments using the SSDD and AirSARship datasets, we illustrate the effectiveness of our suggested model. The SSDD dataset ablation studies of FPN+PAN and attention mechanism modules have shown that each of them can enhance ship detection performance, and the combination of both can increase detection outcomes.

This section compares various CNN-based algorithms, such as Yolov5-X, Faster-RCNN (Lin et al., 2017), SSD (Liu et al., 2016), and R2CNN (Jiang et al., 2017), as well as current methods for object recognition in SAR images using large-scale scene SAR images. To validate the good migration ability of the suggested model, the actual ship identification in two more SAR large scene images has been carried out. Figure 13 displays the areas covered by the two SAR large scene images acquired by the Chinese satellite GF3. These two SAR images were selected because they both lie along globally important routes named Malacca strait ( Table 5 ). The VV polarization SAR images from Table 5 are considered because ships often exhibit higher backscattering values in VV polarization (Torres et al., 2012). Due to the restricted GPU memory, they are divided into 800 x 800 (Zhang et al., 2020a) and followed by (Xian et al., 2019) 1000 x 1000 of small sub images before being used for training and testing. SAR ships are ultimately added to the suggested model in order to actually detect them. The outcomes of the sub- image’s detection are then added to the original SAR large scene image ( Figure 14 ). The detection accuracy and speed of various approaches are assessed, respectively, using the AP and FPS ( Table 6 ).

The two SAR large scene images are used to depict the results of SAR ship detection using the current model and other CNN-based methods. The current model is able to successfully detect the majority of ships, demonstrating its strong migration application capabilities in ocean surveillance ( Figures 15A, B ). Features and a clear environment come first. In particular, just a few ships in the big panorama were missed by all of these convolutional neural network -based approaches, which are indicated by pink circles in Figures 15A, B . The identification outcomes on the second SAR large scene image are displayed, and the identification results of various CNN-based approaches on the SAR large scene images are demonstrated on the left side of Figures 15A, B . Figures 15A, B ’s right side enlarges and displays two particular regions designated with brown and blue rectangles. The comparison of the detection outcomes of various model performances on SAR large scene images is presented in Figures 15A, B . Offshore ships make up the majority of this SAR large-scene image. Inshore scenes have substantial clutter, which could cause false alarms. The results of another CNN-based approach suggested in this study show that there are few false alarms and missed targets in the offshore scenes, and the false alarms are repressed in the inshore scenes as well. However, the proposed approach does not have false alarms or miss target detection, which is one of the motivating factors for this paper. Figures 15A, B illustrates the visualization of the SAR ship identification performance of the suggested model on two SAR large scene images acquired from the Chinese GF3 satellite and employed in the current study. The suggested model can successfully detect the majority of ships, which demonstrates its good migration application capability in ocean surveillance, it can be deduced from Figures 15A, B . In conclusion, the identification outcomes on the two SAR large scene images at various resolutions show that the present system identifies multi size ships with competitive outcomes and has a strong generalization capacity in comparison to the various CNN-based approaches created for identification object in SAR images. It demonstrates that the approach now under study can adapt to SAR images from various sources more effectively.

Table 7 illustrates the ablation research of YOLOv5 removal and installation of the FPN+PAN module. Table 7 , “✘“ denotes YOLOv5 without the FPN+PAN module, while “✔“ denotes YOLOv5 with the FPN+PAN module (i.e., our suggested model). Experiments were carried out in offshore and inshore scenes respectively, to evaluate the identification achievement of the suggested approach for offshore and inshore scenes. In Table 7 , the identification effectiveness of the model is approximately similar, with little modification in identification indicator, as a result of the relatively simple background of offshore ship and little interference, due to the identification achievement has been outstanding in the simple background, and only limited enhancement can be obtained. In contrast to offshore ships, inshore ships have a more complicated backdrop clutter. Moreover, wharfs as well as other structures on the shore significantly undermine detections, and SAR ship detection effectiveness generally declines. However, the detection performance is significantly enhanced by improving the feature representation capacity and minimizing the aliasing effect of fusion features. In Table 7 , by installation of FPN +PAN modules, the result of p, r and AP rate of the model are enhanced by about 2.9%, 4.6% and, 5.5% respectively for offshore ship detection, while for inshore ship detection the p, r and AP rate of the model are enhanced by about 6.08%, 15% and, 8.07% respectively.

Table 8 presented the ablation research of YOLOv5 removal and installation of the attention mechanism module. Table 8 , “✘“ denotes YOLOv5 without the attention mechanism module, “✔“ denotes YOLOv5 with the attention mechanism module (i.e., our suggested model). The installation of attention mechanism module to our model, the detection performance is significantly enhanced by improving the feature representation capacity and minimizing the aliasing effect of fusion features. In Table 8 , through installation of attention mechanism modules, as a result the p, r and AP rate of the model are enhanced by about 7.22%, 1.36% and, 5.31% respectively for offshore ship detection, while for inshore ship detection the p, r and AP rate of the model are enhanced by about 1.03%, 10.97% and, 10.67% respectively.