A camera-based method and system for iceberg instance segmentation, recognition, and detection.

Abstract

This invention discloses a camera-based iceberg instance segmentation, recognition, and detection method and system. The method includes the following steps: acquiring polar sea image data; inputting the image data into a first-stage target detection network to output bounding box prediction results for sea ice targets; converting the output bounding box prediction results into location cues and inputting them into a second-stage cue-driven segmentation network; the cue-driven segmentation network performing pixel-level fine-grained segmentation of sea ice targets in the image based on the received location cues, and outputting sea ice instance masks. This invention effectively alleviates the information loss problem during feature compression through a progressive channel reduction module, improving the recall rate and localization accuracy of the detection network for adhered and small iceberg targets. This invention adopts a two-stage architecture of "detection-led, cue-driven segmentation," fully utilizing the strong localization information of the detection boxes to guide segmentation, achieving more accurate instance-level segmentation boundaries.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and more specifically, to a camera-based iceberg instance segmentation, recognition, and detection method and system. Background Technology

[0002] Accurate sea ice detection is crucial for ensuring ship safety during polar voyages. Existing sea ice identification technologies mainly fall into two categories: First, semantic segmentation-based methods, while capable of distinguishing ice and water pixels, often fail to differentiate between interlocking ice blocks in densely populated sea ice areas, leading to the failure of subsequent dynamic parameter estimations. Second, traditional instance segmentation methods (such as Mask R-CNN), which often struggle to balance detection accuracy and edge segmentation quality, and exhibit poor robustness under complex polar lighting conditions (such as strong reflections and ship shadows).

[0003] In existing technologies, real-time detection models such as RT-DETR typically employ simple 1×1 convolutions when compressing the high-dimensional features of the backbone network (e.g., 2048 channels) to the encoder input dimension (e.g., 256 channels). This drastic dimensionality reduction leads to a significant loss of deep semantic information, limiting the detection capability for small or blurry sea ice targets. Furthermore, while general segmentation models (e.g., SAM2) have strong generalization capabilities, they are prone to misclassification when faced with shadowy regions that are similar in color to seawater.

[0004] Therefore, there is an urgent need for an iceberg identification method that can achieve high-precision detection, fine instance segmentation, and strong robustness against the special lighting and shadow interference in polar regions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a camera-based iceberg instance segmentation, recognition and detection method and system. Through an innovative two-stage architecture, combined with progressive feature dimensionality reduction and physical-driven data augmentation, high-precision sea ice individual recognition and segmentation are achieved.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A camera-based iceberg instance segmentation, recognition, and detection method includes the following steps:

[0008] Step S1: Acquire image data of polar sea areas;

[0009] Step S2: Input the image data into the first-stage target detection network and output the bounding box prediction results of the sea ice target;

[0010] Step S3: Convert the output bounding box prediction results into location cues and input them into the cue-driven segmentation network of the second stage;

[0011] Step S4: The prompt-driven segmentation network performs pixel-level fine-grained segmentation of sea ice targets in the image based on the received location prompts, and outputs a sea ice instance mask.

[0012] Furthermore, in step S2, the target detection network includes a backbone network, a progressive channel reduction module, and a hybrid encoder. The backbone network extracts multi-scale feature maps from the input image, and after feature dimension compression by the progressive channel reduction module, it is input to the hybrid encoder.

[0013] Furthermore, the progressive channel reduction module adopts a multi-level stepwise dimensionality reduction strategy to gradually reduce high-dimensional features to low dimensions. Each step of the dimensionality reduction operation includes depthwise separable convolution and channel attention mechanisms.

[0014] Furthermore, the progressive channel reduction module is constructed as follows:

[0015] Let the input feature map be The channel dimension is The output target dimension is Define the dimensionality reduction step size sequence ,in, , ;

[0016] For the Step-by-step dimensionality reduction, input features The feature transformation process is as follows:

[0017]

[0018] In the formula, Indicates the first Input features after dimensionality reduction; This represents a depthwise separable convolution operation; This indicates compression and excitation operations.

[0019] Furthermore, in step S4, the cue-driven segmentation network includes an image encoder, a cue encoder, and a mask decoder. The cue encoder receives the generated bounding box coordinates and encodes them into a sparse embedding vector, guiding the mask decoder to focus on the sea ice region in the feature embedding.

[0020] Furthermore, during the model training phase, a label-guided shadow enhancement strategy is introduced. By generating and overlaying a semi-transparent shadow mask at the geometric center of the labeled sea ice instances, the model learns a robust representation of shadows.

[0021] Furthermore, the implementation steps of the label-guided shadow enhancement strategy include:

[0022] Step A1: Read the ground truth mask of the training samples and calculate the geometric center of each sea ice instance. ;

[0023] Step A2, with the geometric center As a baseline, apply a random offset To determine the center position of the shadow;

[0024] Step A3: Generate a semi-transparent black mask. And blended and overlaid onto the original image using a transparency factor. Generate enhanced images ;

[0025] Step A4: Apply shadow enhancement only to the annotated area of ​​the sea ice instance, leaving the background area unchanged.

[0026] Furthermore, in step A3, the image is enhanced. Calculated and generated using the following formula:

[0027]

[0028] In the formula, Transparency coefficient; This is a mask for binarizing shadow regions.

[0029] The present invention also provides a camera-based iceberg instance segmentation, recognition, and detection system, comprising:

[0030] The image acquisition module is used to acquire real-time sea condition images via a shipborne camera.

[0031] A processor is used to implement the steps of the camera-based iceberg instance segmentation, recognition, and detection method described above when executing a computer program;

[0032] The navigation assistance module receives the sea ice instance mask output by the processor and calculates the passable area based on the sea ice distribution.

[0033] The beneficial effects of this invention are:

[0034] 1. This invention effectively alleviates the problem of information loss during feature compression by using a progressive channel reduction module, thereby improving the recall rate and positioning accuracy of the detection network for adhered and tiny iceberg targets.

[0035] 2. This invention adopts a two-stage architecture of "detection-guided segmentation" to make full use of the strong localization information of the detection box to guide segmentation and achieve more accurate instance-level segmentation boundaries.

[0036] 3. This invention significantly improves the robustness of the model to changes in lighting and shadow occlusion, and reduces the missegmentation rate by adopting a label-guided shadow enhancement training strategy. Attached Figure Description

[0037] Figure 1 This is a flowchart of a camera-based iceberg instance segmentation, recognition, and detection method in this embodiment;

[0038] Figure 2 This is a structural framework diagram of the progressive channel reduction module in this embodiment;

[0039] Figure 3 This is a structural framework diagram of the camera-based iceberg instance segmentation, recognition, and detection system in this embodiment;

[0040] Figure 4 This is a schematic diagram illustrating one effect of label-guided shadow enhancement in this embodiment.

[0041] Figure labels: Image acquisition module 1, processor 2, navigation assistance module 3. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] A camera-based iceberg instance segmentation and detection method is proposed. This method constructs a two-stage processing architecture. The first stage is an improved object detection network (named IceDETR) used to generate high-precision sea ice bounding boxes. The second stage is a cue-driven segmentation network (based on the SAM2 model), which uses the bounding boxes from the first stage as cues to generate a fine-grained mask. To address the feature loss problem, this invention designs a progressive channel reduction module. Unlike traditional one-time dimensionality reduction, this module combines depthwise separable convolution with squeeze-excited (SE) attention mechanisms to gradually reduce the number of channels in multiple steps (e.g., 2048→1024→512→256), maximizing the preservation of semantic features beneficial for detection. To address the shadow interference problem, this invention proposes a label-guided shadow enhancement strategy. This strategy simulates the optical properties of physical shadows (brightness attenuation rather than structural change) and forces the model to learn shadow invariance on labeled objects during the training phase, significantly reducing the probability of misidentifying shadows as seawater.

[0044] It should be noted that, in the context of shipborne visual near-field perception, "iceberg instance" is used to refer to floating ice, broken ice, adhered sea ice, and ice targets of different scales at varying distances that need to be identified in front of the ship, which is equivalent to "sea ice target".

[0045] Specifically, such as Figures 1-2As shown, the method includes the following steps:

[0046] Step S1: Acquire polar sea image data collected by the ship's camera. The image data includes sea ice, seawater, and background information.

[0047] Step S2 involves inputting image data into the first-stage target detection network, which includes a backbone network, a progressive channel reduction module, and a hybrid encoder. The backbone network extracts multi-scale feature maps from the image. These feature maps undergo progressive channel reduction for dimensionality reduction, and the processed features are fused by the hybrid encoder. Finally, the decoder outputs a series of bounding boxes for sea ice targets. These bounding boxes serve as location cues for the second-stage cue-driven segmentation network. This process reduces computation while preserving, to the maximum extent possible, the deep semantic information crucial for detecting small, adherent targets.

[0048] See Figure 2 The progressive channel reduction module replaces the traditional 1×1 convolution operation used in convolutional neural networks for dimensionality reduction. This module employs a multi-level stepwise dimensionality reduction strategy to progressively reduce high-dimensional features to low-dimensional ones. Each step of the dimensionality reduction operation includes depthwise separable convolution and a channel attention mechanism (Squeeze-and-Excitation, SEBlock) to address the information bottleneck problem when compressing high-dimensional features (e.g., 2048 channels) to low-dimensional features (e.g., 256 channels).

[0049] Let the input feature map be The channel dimension is The output target dimension is The progressive channel reduction module decomposes the dimensionality into A gradual step, dimensionality reduction step size sequence ,in, , ;

[0050] For the Step-by-step dimensionality reduction, input features After a "convolution-calibration" feature transformation process, which is defined by the following formula:

[0051]

[0052] In the formula, Indicates the first Input features after dimensionality reduction; ; This indicates a depthwise separable convolution operation, used to reduce computation while preserving spatial information; This indicates squeezing and excitation operations used for adaptive recalibration of channel weights.

[0053] For depthwise separable convolution operations Specifically, this includes depthwise convolution (DWConv) and pointwise convolution (PWConv). The operation first performs a 3×3 depthwise convolution (DWConv) to independently spatially filter each input channel to maintain spatial resolution; then, a 1×1 pointwise convolution (PWConv) is performed to linearly combine the features from different channels, achieving channel blending and dimensionality reduction to obtain intermediate features. This operation reduces computational complexity while preserving spatial details; its calculation is as follows:

[0054]

[0055] In the formula, It is a 3×3 depthwise convolution; It is a 1×1 pointwise convolution.

[0056] For squeezing and excitation operations Specifically, this operation applies a channel attention mechanism (SEBlock) to the intermediate features of the depthwise separable convolution output. First, spatial information is compressed into channel descriptors through global average pooling. Subsequently, channel weights are learned through two fully connected layers. The calculation is as follows:

[0057]

[0058] In the formula, For the Sigmoid function; It is the ReLU activation function; and These are the weight matrices of the two fully connected layers; the final output features are... This achieves adaptive weighting at the channel dimension.

[0059] Through multiple iterations of the above "convolution-calibration" steps, the feature channels are smoothly and faithfully reduced to the target dimension. This provides richer feature representations for subsequent encoding and detection.

[0060] Step S3: Convert the bounding box prediction results output in step S2 into location cues and input them into the cue-driven segmentation network in the second stage.

[0061] Step S4: The prompt-driven segmentation network performs pixel-level fine-grained segmentation of sea ice targets in the image based on the received location prompts, and outputs a sea ice instance mask.

[0062] The cue-driven segmentation network employs a foundation model based on the Transformer architecture. This network includes an image encoder, a cue encoder, and a mask decoder. The cue encoder receives the bounding box coordinates generated in step S2. The data is encoded into sparse embedding vectors that guide the mask decoder to focus on the sea ice region corresponding to the location cue in the overall feature embedding of the image. The mask decoder performs pixel-level classification and outputs a fine-grained instance segmentation mask for each sea ice region.

[0063] Furthermore, to improve the model's performance under complex polar lighting conditions, a label-guided shadow augmentation strategy is introduced during the model training phase of this method. This strategy generates and overlays a semi-transparent shadow mask at the geometric center of the labeled sea ice instance to enhance the model's robustness to lighting changes and shadow occlusion.

[0064] Specifically, the data is augmented during the training phase as follows:

[0065] Step A1: For each sea ice instance in the training samples, a true mask is created. Calculate its geometric center ;

[0066] Step A2, with the geometric center As a baseline, apply a random offset To determine the center position of the shadow;

[0067] Step A3: Generate a semi-transparent black mask. And through transparency coefficient The blending method is overlaid onto the original image Above, generate enhanced image Its pixel-level calculation formula is:

[0068]

[0069] In the formula, This is the transparency coefficient, with a value range of [0.3, 0.7]. This is a mask for binarizing shadow regions.

[0070] Step A4: Apply shadow enhancement only to the marked area of ​​the sea ice instance, while leaving the background seawater and sky areas unchanged.

[0071] This "label-guided" strategy ensures that the model, during learning, associates "sea ice with shadows" with "sea ice" itself, rather than learning the shadow as a new background or object. This effectively improves the model's segmentation accuracy under real shadow interference. Figure 4 As shown.

[0072] This embodiment also provides a camera-based iceberg instance segmentation, recognition, and detection system, such as... Figure 3 As shown, it includes: an image acquisition module 1, a processor 2, and a navigation assistance module 3; wherein, the image acquisition module 1 is used to acquire real-time sea state images through a shipborne camera; the processor 2 is used to implement the steps of the above-mentioned camera-based iceberg instance segmentation, recognition, and detection method when executing a computer program; the navigation assistance module 3 is used to receive the sea ice instance mask output by the processor 2 and calculate the passable area according to the sea ice distribution.

[0073] The following section provides a more detailed explanation of the invention, including the training data construction process, network structure parameters, shipborne online inference scenario, and performance verification examples.

[0074] Example 1: Training Data Construction and Labeling Process

[0075] This embodiment selects continuous images captured by a shipborne camera during polar navigation as the original sample. The camera can be mounted forward of the ship's bridge or above the bow, with a field of view covering the sea surface in front of the ship, near-field floating ice, and distant ice zones. The preferred resolution of the original images is 2560×1440, and the acquisition environment includes complex conditions such as sunny days, cloudy days, foggy days, backlighting, solar reflection, and ship shadows. After deduplication, removal of severely out-of-focus frames, and invalid completely black frames from the original images, a sample set of 841 images is established, of which 672 images are used as the training set and 169 images are used as the test set. For visible sea ice targets in each image, both bounding boxes and instance masks are simultaneously labeled; for ice zones far from the ship but still relevant to the navigation environment, they can be further labeled separately as the farIce category, thus taking into account the perception needs of both near-field sea ice and far-field ice zones.

[0076] Before training, image samples undergo preprocessing including normalization, random flipping, scale perturbation, and color perturbation. Combined with the label-guided shadow enhancement strategy (described later), this improves the model's generalization ability under different weather and lighting conditions. During training, the AdamW optimizer can be used, with the global learning rate of the detection network set to [value missing]. The backbone network learning rate is set to The weight decay is set to The gradient clipping upper limit is set to 0.5. Training is optimally performed for 400 epochs, employing a cosine annealing hot restart learning rate strategy and exponential moving average parameter updates to improve convergence stability and final detection accuracy.

[0077] Example 2: Specific Structure of the First-Stage Target Detection Network

[0078] The first-stage object detection network in this embodiment is based on the RT-DETR architecture, including a backbone network, a progressive channel reduction module, a hybrid encoder, and a decoder. The backbone network outputs feature maps at least three scales, with channel numbers of 512, 1024, and 2048, respectively. For deep features with 2048 channels, instead of directly using 1×1 convolutions to reduce to 256 channels, dimensionality reduction is performed progressively along the path 2048→1024→512→256. For features with 1024 channels, dimensionality reduction can be performed along the path 1024→512→256. For features with 512 channels, dimensionality reduction can be performed directly to 256 channels, or after passing through a convolutional-calibration unit. Each dimensionality reduction step includes depthwise separable convolution and a SE channel attention mechanism. Depthwise convolution is used to preserve spatial details, pointwise convolution is used to achieve channel blending and dimensionality reduction, and the SE mechanism is used to recalibrate the channels of the intermediate features after dimensionality reduction. The multi-scale feature input after the above processing is used to fuse global and local information, and then the decoder outputs sea ice candidate boxes and their category confidence scores.

[0079] In practical deployment, confidence thresholds and non-maximum suppression strategies can be set to filter low-quality bounding boxes. For example, candidate boxes with confidence scores below a preset threshold can be directly eliminated, while the highest-scoring candidate box with high overlap pointing to the same sea ice target can be retained. The filtered bounding boxes serve as location cue inputs for the second-stage cue segmentation network.

[0080] Example 3: Second-stage prompt-driven segmentation and shadow enhancement process

[0081] In this embodiment, the second-stage network employs a cue-driven segmentation model, preferably SAM2 or an equivalent basic segmentation model. For each sea ice bounding box output from the first stage... The original image is converted into a location cue and encoded into a sparse cue vector by a cue encoder. An image encoder generates dense image features from the original image. Guided by the location cue, a mask decoder performs fine-grained pixel-level segmentation of the corresponding bounding box region, outputting at least one candidate instance mask. For multiple candidate masks, the final output mask can be selected based on mask quality score, matching degree with the bounding box, or mask area continuity. Preferably, a consistency check is performed between the mask boundary and the original bounding box; if the mask exceeds the bounding box by too much or its area is too small, it is discarded or the result is reverted to the detection box result.

[0082] To enhance the network's robustness to hull shadows, backlit dark areas, and local occlusion, label-guided shadow enhancement is implemented for each labeled instance during the training phase. Specifically, the geometric center is first calculated based on the instance's binary mask. Subsequently, a random offset was introduced near the geometric center. The offset is preferably no more than 20% of the width and height of the target bounding box; then a local shadow mask is generated according to the target scale, the mask radius is preferably 40%-100% of the target equivalent size, and the transparency coefficient is... The optimal range for the shadow value is 0.3–0.7, and the shadow enhancement is applied to the training samples with a probability of 0.9. Shadow masking only applies within the instance's labeled area; the background seawater and sky areas remain unchanged. This allows the model to learn the discrimination rule that "sea ice with shadows is still sea ice," rather than treating the shadow as a new background category. With this enhancement strategy, the second-stage model significantly reduces the number of missed segments in the hull shadow-covered areas.

[0083] Example 4: Shipborne Online Application Scenarios

[0084] This embodiment deploys the invention in the forward perception system of an icebreaker or polar research vessel. The image acquisition module periodically acquires sea state images ahead of the ship and sends them to a processor installed on the ship's computing platform. The processor first performs a first-stage target detection, obtaining bounding boxes of multiple sea ice targets in the current frame; then, each bounding box is converted into a second-stage prompt input, obtaining the corresponding instance mask results; finally, the navigation assistance module identifies navigable waters ahead of the ship, potential collision risk areas, and dense ice areas requiring deceleration and avoidance based on the position, area, spacing, and relative relationship of each instance mask to the ship's bow direction. For continuous video streams, instance results from adjacent frames can also be correlated to obtain the sea ice drift direction and movement trend, providing continuous environmental perception results for the pilot or automatic navigation assistance module.

[0085] Example 5: Effect Verification Example

[0086] The present invention was validated using the dataset described in Example 1. Under the same hardware and software environment, the present invention was compared with Mask R-CNN, Cascade Mask R-CNN, Mask2Former, and YOLO series instance segmentation methods. The present invention showed superior results in both object detection and instance segmentation tasks. Specifically, the average precision (AP) of the detection module reached 49.1%, and the intersection-over-union (IoU) of the second-stage segmentation module under bounding box cues reached approximately 80%. In end-to-end instance segmentation comparisons, the detection mAP of the present invention was approximately 0.491, and the segmentation mAP was approximately 0.400, achieving more stable results in scenarios involving adhered sea ice, small-scale targets, and shadow occlusion. The above validation demonstrates that the present invention not only improves the recall and localization accuracy of sea ice targets but also enhances the segmentation boundary quality in complex lighting scenarios, possessing high engineering practical value.

[0087] Furthermore, to present the validation results in a tabular format, the validation can use the following dataset, evaluation metrics, and comparison results. The sample set contains 841 polar sea images captured by a shipborne camera, with an image resolution of 2560×1440. 672 images are used as the training set, and 169 as the test set. The evaluation categories include near-field sea ice targets and far-field ice zone targets. mAP, mAP50, and mAP75 are used as evaluation metrics for both target detection and instance segmentation tasks.

[0088] The target detection performance comparison results are shown in Table 1, where mAP is calculated based on the average accuracy within the IoU threshold range of 0.5 to 0.95.

[0089] Table 1 Comparison of Target Detection Performance

[0090]

[0091] The comparison results of instance segmentation performance are shown in Table 2. As can be seen from the results in Table 2, the present invention outperforms the comparison methods in terms of segmentation mAP, mAP50, and mAP75, indicating that while achieving higher positioning accuracy, it can obtain instance masks that better fit the contours of real sea ice.

[0092] Table 2 Instance segmentation performance comparison

[0093]

[0094] Furthermore, to verify the effectiveness of the progressive channel reduction module in the first-stage target detection network, the basic RT-DETR, the direct dimensionality reduction form IceDETR-direct, and the progressive dimensionality reduction form IceDETR-step2 were compared and ablated. The results are shown in Table 3.

[0095] Table 3. Comparison of ablation methods using progressive channel reduction modules

[0096]

[0097] The ablation results above show that the direct dimensionality reduction method offers a certain improvement over the basic RT-DETR. Furthermore, the progressive dimensionality reduction method increases the detection mAP from 0.479 to 0.491 and the mAP50 from 0.685 to 0.691, indicating that multi-level dimensionality reduction can reduce semantic information loss during high-dimensional feature compression. Although this structure introduces a small increase in parameters and a slight change in speed, its improved detection accuracy is more beneficial for shipborne sea ice sensing.

[0098] Furthermore, when visually comparing images of adhered sea ice, small-scale floating ice partially exposed above the water, sunlight reflection, low exposure in backlight, and ship shadow occlusion, traditional instance segmentation methods are prone to problems such as missed detection of far-field sea ice, false detection of sea surface reflection, unclear boundaries of adhered targets, or misclassification of shadowed areas as background. This invention can more stably separate adhered sea ice instances and maintain a relatively continuous mask boundary in shadow-covered areas. For the second-stage cue-driven segmentation network, after adopting label-guided shadow enhancement, the missed segmentation phenomenon in shadow-covered sea ice areas is significantly reduced; under box cue conditions, its intersection-over-union (IoU) can reach approximately 80%.

[0099] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A camera-based iceberg instance segmentation and recognition detection method, characterized in that, Includes the following steps: Step S1: Acquire image data of polar sea areas; Step S2: Input the image data into the first-stage target detection network and output the bounding box prediction results of the sea ice target; Step S3: Convert the output bounding box prediction results into location cues and input them into the cue-driven segmentation network of the second stage; Step S4: The prompt-driven segmentation network performs pixel-level fine-grained segmentation of sea ice targets in the image based on the received location prompts, and outputs a sea ice instance mask; During the model training phase, a label-guided shadow enhancement strategy is introduced, which generates and overlays a semi-transparent shadow mask at the geometric center of the labeled sea ice instances, enabling the model to learn a robust representation of shadows.

2. The camera-based iceberg instance segmentation and detection method according to claim 1, wherein, In step S2, the target detection network includes a backbone network, a progressive channel reduction module, and a hybrid encoder. The backbone network extracts multi-scale feature maps from the input image, and after feature dimension compression by the progressive channel reduction module, it is input to the hybrid encoder.

3. The camera-based iceberg instance segmentation and detection method according to claim 2, wherein, The progressive channel reduction module employs a multi-level stepwise dimensionality reduction strategy to gradually reduce high-dimensional features to low-dimensional ones. Each step of the dimensionality reduction operation includes depthwise separable convolution and channel attention mechanisms.

4. The camera-based iceberg instance segmentation and recognition detection method according to claim 3, wherein, The progressive channel reduction module is constructed as follows: Let the input feature map be The channel dimension is The output target dimension is Define the dimensionality reduction step size sequence ,in, , ; For the Step-by-step dimensionality reduction, input features The feature transformation process is as follows: In the formula, Indicates the first Input features after dimensionality reduction; This represents a depthwise separable convolution operation; This indicates compression and excitation operations.

5. The camera-based iceberg instance segmentation and detection method according to claim 1, wherein, In step S4, the cue-driven segmentation network includes an image encoder, a cue encoder, and a mask decoder. The cue encoder receives the generated bounding box coordinates and encodes them into a sparse embedding vector, guiding the mask decoder to focus on the sea ice region in the feature embedding.

6. The camera-based iceberg instance segmentation and detection method according to claim 1, wherein, The implementation steps of a label-guided shadow enhancement strategy include: Step Al, read the ground truth mask of the training sample and compute the geometric center of each sea ice instance ; Step A2, with the geometric center as a reference, apply random offset to determine the center position of the shadow; Step A3: Generate a semi-transparent black mask. And blended and overlaid onto the original image using a transparency factor. Generate enhanced images ; Step A4: Apply shadow enhancement only to the annotated area of ​​the sea ice instance, leaving the background area unchanged.

7. The camera-based iceberg instance segmentation and detection method according to claim 6, wherein, In step A3, the enhanced image is calculated generated according to the following formula: In the formula, is the transparency coefficient; is the binary shadow region mask.

8. A camera-based iceberg instance segmentation and recognition detection system, characterized in that, include: The image acquisition module is used to acquire real-time sea condition images via a shipborne camera. A processor, configured to implement the steps of the camera-based iceberg instance segmentation, recognition, and detection method as described in any one of claims 1-7 when executing a computer program; The navigation assistance module receives the sea ice instance mask output by the processor and calculates the passable area based on the sea ice distribution.

Images