Modular underwater cleaning and monitoring integrated ROV method and system

Abstract

The invention provides a modularly designed underwater cleaning and monitoring integrated ROV method and system, and belongs to the technical field of underwater robot control. Comprising a body, the body comprises a plurality of function modules, the function modules have independent functions, and the adjacent function modules are detachably connected; the propellers are arranged at different positions of the body and used for driving the body to move underwater in all directions; the multi-source sensor array is integrated on the body and is used for collecting an underwater environment, obtaining water body parameters and images and realizing multi-dimensional perception; wherein the body is internally provided with a lightweight multi-scale image enhancement network and an SC-YOLO algorithm, and the lightweight multi-scale image enhancement network is used for performing feature map extraction on an image acquired by the multi-source sensor array to generate an enhanced image; and the SC-YOLO algorithm takes the enhanced image as input, carries out underwater garbage recognition and feeds back the recognition result to a functional module of the body, and executes automatic path-finding and cleaning operation.

Description

Technical Field

[0001] This invention relates to the field of underwater robot control technology, and in particular to a modular design of an integrated underwater cleaning and monitoring ROV method and system. Background Technology

[0002] Remotely operated underwater vehicles (ROVs), as core equipment for underwater operations, have been gradually applied to the monitoring and cleaning of marine ranches. However, existing ROVs have many shortcomings: such as difficulty in disassembly and maintenance, poor flexibility in functional expansion; insufficient rationality in the distribution of power layout, with some ROVs having flawed thruster layout design, resulting in poor motion flexibility and inability to achieve precise attitude control in all directions, affecting operational efficiency; underwater imaging technology is difficult to adapt to complex aquatic environments, with prominent issues of image blurring and color distortion; target detection algorithms have insufficient accuracy in identifying low-resolution, multi-scale underwater debris targets, resulting in a high false negative rate, meaning that the cleaning of underwater debris, such as microplastics and discarded fishing nets, still relies mainly on manual retrieval, which is inefficient and poses significant safety hazards.

[0003] Therefore, it is essential to provide a modularly designed integrated underwater cleaning and monitoring ROV method and system that enables omnidirectional flexible movement through a rationally arranged thruster, adopts a modular structure to achieve rapid loading, unloading and expansion of the equipment, improves imaging quality in complex aquatic environments, accurately identifies underwater debris targets, and realizes an integrated solution for underwater cleaning and monitoring. Summary of the Invention

[0004] In view of this, the present invention proposes a modular design for an integrated underwater cleaning and monitoring ROV method and system that is simple to maintain, highly scalable, more suitable for complex aquatic environments, and has a good recognition effect on underwater low-resolution and multi-size garbage targets.

[0005] On the one hand, this invention provides a modularly designed integrated underwater cleaning and monitoring ROV method, comprising the following steps: S1: An integrated ROV system for underwater cleaning and monitoring with a modular design, including the main body and a multi-source sensor array; the multi-source sensor array specifically includes a camera, a multi-beam sonar, a gyroscope, a depth sensor and a turbidity sensor, which respectively acquire underwater images, underwater obstacles, the attitude of the main body, the current depth and the turbidity of the water and feed them back to the main body; S2: The lightweight multi-scale image enhancement network built into the ontology performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images to enhance input characteristics and improve the overall image effect. Then, the adaptive enhancement module performs multi-scale feature modulation to achieve adaptive feature enhancement. Finally, the multi-dimensional color enhancement module enhances color information from three dimensions: channel, space, and depth, to improve color fidelity and detail sharpness, and outputs the enhanced feature map. S3: Use the improved SC-YOLO algorithm architecture to perform underwater debris target detection and output the target localization results; S4: Based on the positioning results and the body attitude obtained by the gyroscope, the dynamic path planning model improved by the A* algorithm is used to plan the path of the body, so as to realize the obstacle avoidance path planning of the body underwater. S5: Upon reaching the underwater debris site, the main unit drives the cleaning module to clean up the underwater debris.

[0006] Based on the above technical solutions, preferably, the lightweight multi-scale image enhancement network built into the ontology described in step S2 performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images, enhancing input characteristics and improving the overall image effect. This is achieved by using a multi-scale feature refinement module to process the image. This module includes a head, depthwise separable convolutional units, multi-branch dilated depthwise convolutional units, feature fusion units, and a tail, arranged sequentially. The head includes sequentially arranged 1×1 convolutional layers and batch normalization (BN) layers. The 1×1 convolutional layers are used for dimensionality reduction to adjust the original number of channels. The BN layers stabilize the training process and introduce non-linearity through the GELU activation function. The depthwise separable convolutional units are 5×5, using... A 5×5 convolutional kernel performs convolution operations on each channel of the features output from the head section, processing spatial information without mixing between channels. The features processed by the depthwise separable convolutional unit are then fed into a multi-branch dilated depth convolutional unit. The multi-branch dilated depth convolutional unit includes several parallel branches, each branch containing pairs of dilated convolutional layers. The convolutional kernels of the dilated convolutional layers in each branch have the same depth, but different receptive fields. The feature fusion unit is used to fuse the feature maps output from several branches, so that the fused feature map contains information at different scales simultaneously. The tail section has the same structure as the head section, including a 1×1 convolutional layer, a batch normalization (BN) layer, and a GELU activation function, which is used to map the fused feature map back to the original number of channels, completing multi-scale feature refinement.

[0007] Preferably, the adaptive enhancement of features achieved by performing multi-scale feature modulation through the adaptive enhancement module in step S2 is the result of the multi-scale feature refinement module. After inserting it into each convolutional layer of the encoder of the UNet network, features are extracted through three parallel paths. The first path uses 1×1 convolution and dilated convolution to capture image details, the second path uses 3×3 convolution to extract the feature contours of the target, and the third path uses 5×5 convolution and global average pooling layer to capture the overall structure of the image and environmental background information. The outputs of the three parallel paths are then fused by weighted summation, with the weights dynamically allocated by the attention mechanism. The output is an integrated feature map containing shallow detail features. Then, at the decoder end of the UNet network, the original feature fusion module in the decoder structure is replaced with an adaptive enhancement module. The adaptive enhancement module includes a main path and a gated branch. The main path includes a first Conv layer, a first adder, a CMBA module, a second adder, and a second Conv layer arranged sequentially. The first Conv layer performs a 1×1 convolution on the input fin for channel number reduction adjustment. The first adder adds the dimensionality-reduced output of the first Conv layer to the input fin element-wise, and then sends the output to the CMBA module. The CMBA module is an attention module that receives the features input from the first adder and outputs an enhanced feature that has been recalibrated for both channel and spatial dimensions. The second adder adds the output of the CMBA module to the input fin again element-wise. The second Conv layer performs a 1×1 convolution on the features received from the second adder output. Convolution maps features back to the same number of channels as the input fin. The gated branch includes sequentially arranged BN layers, PreLU layers, Pooling layers, first Linear layers, ReLU layers, second Linear layers, Sigmoid layers, and multipliers. The BN layer is used for batch normalization of the input features. The PreLU layer is an activation function with learning parameters used to introduce non-linearity. The Pooling layer is used for global average pooling, compressing the entire feature map into a scalar. The first Linear layer is used to perform dimensionality reduction and non-linear transformation on the pooled channel vectors. The ReLU layer is used to introduce non-linearity after the first Linear layer. The second Linear layer is used to restore the feature dimension to the original number of channels. The Sigmoid layer is used to map the output of the second Linear layer to the (0,1) interval to generate an adaptive weight vector δ. The multiplier multiplies the output of the CMBA module with the adaptive weight vector δ generated by the gated branch element-wise to obtain the enhanced deep structural features, which are used as the output of the adaptive enhancement module. The decoder uses dynamic convolution kernel adjustments to ensure scale matching between the encoder's downsampled features and the decoder's upsampled features.

[0008] Further preferred methods include lightweight modifications and dynamic parameter calibration of the UNet network. Lightweight modifications involve removing redundant convolutional channels from the encoder. Dynamic parameter calibration involves dynamically adjusting the weight allocation strategy between the shallow detail features transmitted by the encoder and the deep structural features generated by the decoder, based on the turbidity values ​​from the turbidity sensor. Specifically: , NTU It is a unit of turbidity.

[0009] Further preferably, in step S2, the color information is enhanced from three dimensions—channel, space, and depth—by a multi-dimensional color enhancement module to improve color fidelity and detail sharpness, and output an enhanced feature map. This involves adaptively enhancing features in the three dimensions of color, channel, and space using a multi-dimensional color enhancement module. The multi-dimensional color enhancement module includes a multi-scale feature fusion stage, a multi-dimensional attention feature enhancement stage, and a feature reconstruction and residual output stage. The multi-scale feature fusion stage includes a first BatchNorm layer, several parallel convolutional layers, a first channel concatenation layer C1, and a feature transformation sequence step. The first BatchNorm layer is used for batch normalization, using the result of the multi-scale feature refinement module as the current input feature for standardization. Several parallel convolutional layers are used to extract features of different scales from the standardized current input features in parallel, through parallel 1×1 convolutions, 3×3 convolutions, and so on. Convolution and 5×5 convolution are used to simultaneously perceive features of different scales in the image; the first channel concatenation layer C1 is used to concatenate the feature maps output by several parallel convolutional layers along the channel dimension to form a multi-scale feature tensor; the feature transformation sequence includes a first 1×1 Conv layer, a first GELU layer, a second 1×1 Conv layer, and a third adder set sequentially, which are used to reduce the number of channels, introduce nonlinearity, and map the features back to the original number of channels, respectively, and then add them element-wise with the current input features; The multi-dimensional attention feature enhancement stage includes parallel Channel-wise Attention, Spatial-wise Attention, 3D-wise Attention, and a second channel stitching layer C2. Channel-wise Attention dynamically adjusts the gain of each RGB channel based on its attenuation characteristics. Spatial-wise Attention performs differential color correction on the target and background regions of the image. 3D-wise Attention combines depth sensor data to adjust the color compensation coefficient based on the current depth of the image. The second channel stitching layer C2 weights and stitches the outputs of Channel-wise Attention, Spatial-wise Attention, and 3D-wise Attention in the channel dimension, aggregating enhanced features from the channel, spatial, and joint 3D perspectives. The multi-dimensional attention feature enhancement stage also constructs a dynamic water parameter adaptation mechanism to dynamically adjust the weights of each branch of Channel-wise Attention, Spatial-wise Attention, and 3D-wise Attention. The feature reconstruction and residual output stage includes a second BatchNorm layer, a third 1×1 Conv layer, a second GELU layer, a fourth 1×1 Conv layer, and a fourth adder. The second BatchNorm layer is located between the multi-scale feature fusion stage and the multi-dimensional attention feature enhancement stage, and is used for batch normalization again. The third 1×1 Conv layer, the second GELU layer, the fourth 1×1 Conv layer, and the fourth adder are located after the multi-dimensional attention feature enhancement stage. They are used to reduce the dimensionality of the result after the second channel splicing layer C2, introduce nonlinearity again, and perform the final projection to map the features back to the same dimension as the current input features. The fourth adder then adds the feature element-wise to the current input features.

[0010] More preferably, the dynamic adaptation mechanism for water parameters is based on the data collected from water turbidity and the operating depth of the device using turbidity and depth sensors, to determine the following branch weights: 1) Turbidity > 300 NTUFor depths > 30 meters, the weight of Channel-wise Attention is increased to 60%, and the combined weights of Spatial-wise Attention and 3D-wise Attention are increased to 40%. 2) 150 NTU ≤Turbidity≤300 NTU For depths of 10 meters to 30 meters, the weight of Channel-wise Attention is 35%, the weight of Spatial-wise Attention is 35%, and the weight of 3D-wise Attention is 30%. 3) Turbidity <150 NTU For depths <10 meters, the weight of the Spatial-wise Attention branch is 55%; the sum of the weights of the Channel-wise Attention branch and the 3D-wise Attention branch is 50%.

[0011] More preferably, step S3 is as follows: the SC-YOLO algorithm replaces the C2f module in the feature fusion network of the YOLOv8 algorithm architecture with the C2f_SCConv module. The C2f_SCConv module specifically includes an input splitting unit, a C2F basic feature extraction unit, an SCConv self-calibration feature optimization unit, and a feature fusion and aggregation unit. The input splitting unit is used to further split the input feature map into two parallel main feature branches and calibration feature branches after passing it through a continuously set convolutional Conv-batch normalized BN-SiLU activation function. The main feature branch is used to extract global features from the feature map, and the calibration feature branch is used to calibrate local details of the feature map. The ratio of the main feature branch to the calibration feature branch is 1:1. The C2F basic feature extraction unit contains sequentially arranged Bottleneck sub-units. Each Bottleneck sub-unit includes 1×1 convolutional dimensionality reduction, 3×3 convolutional feature extraction, and residual connection structure, which is suitable for underwater garbage targets with a size of 5cm-50cm. The SCConv self-calibrated feature optimization unit performs multi-scale window sliding sampling on the feature map through dynamic field perception adjustment, capturing the edge and texture features of underwater debris targets and suppressing background noise; The feature fusion and aggregation unit fuses the features of the main feature branch and the calibration feature branch through element-wise addition and pointwise convolution, then unifies the feature dimensions through a 1×1 convolutional layer, and finally enhances the non-linear expressive power of the features through the SiLU activation function, outputting the calibrated feature map.

[0012] On the other hand, the present invention provides a modularly designed integrated underwater cleaning and monitoring ROV system, comprising: The main body consists of several functional modules, each with its own independent function, and adjacent functional modules can be detachably connected. Several thrusters are installed at different positions on the main body to drive the main body to move in all directions underwater; A multi-source sensor array, integrated on the main body, is used to collect underwater environment data, obtain water parameters and images, and achieve multi-dimensional perception. The system incorporates the aforementioned modular design of an integrated underwater cleaning and monitoring ROV method. A lightweight multi-scale image enhancement network is used to extract feature maps from images acquired by a multi-source sensor array to generate enhanced images. The SC-YOLO algorithm uses the enhanced images as input to identify underwater debris and feeds it back to the system's functional modules to perform cleaning operations.

[0013] Preferably, the functional modules include a main control module, a power output module, an environmental sensing module, a cleaning module, and an energy storage module. The main control module is communicatively connected to the power output module, the environmental sensing module, and the cleaning module. The energy storage module provides energy to the main control module, the power output module, the environmental sensing module, the cleaning module, and the energy storage module. Each functional module has a cuboid structure, and each side surface is provided with a slider and a groove. Adjacent functional modules directly transmit energy or communicate via an aviation connector. A multi-source sensor array is communicatively connected to the environmental sensing module.

[0014] In a further preferred embodiment, the power output modules are located at each vertex position of the main control module and on one side surface of the main control module. Each power output module is equipped with a corresponding thruster. The thrusters at each vertex position of the main control module are used to drive the body to move in a planar manner underwater, and the thrusters on the side surface of the main control module are used to adjust the depth of the body underwater.

[0015] The modular design of the integrated underwater cleaning and monitoring ROV method and system provided by this invention has the following advantages compared to existing technologies: 1. The modular body structure proposed in this application enables rapid switching between different functional modules and accurate fault diagnosis. The standard interface supports the expansion of different sensors and robotic arms, making it suitable for different underwater tasks. 2. The lightweight multi-scale image enhancement network built into the system performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images, enhances input characteristics and improves the overall image effect, effectively solves the image distortion problem caused by water turbidity and light attenuation, and significantly improves image clarity, color fidelity and detail sharpness, providing high-precision visual support for target detection and ecological monitoring. 3. The improved SC-YOLO algorithm can still maintain a high ability to identify small targets and hidden debris in complex underwater environments, reduce the false detection rate and the missed detection rate, and improve the identification ability and operation efficiency of cleaning. 4. In the underwater path planning of the main body, an improved A* algorithm is adopted, which integrates distance-velocity dual threshold obstacle avoidance logic. The path weight is dynamically adjusted according to the water environment, and the target priority or safety priority mode is adaptively selected. In areas with dense garbage, priority is given to approaching the target, and in narrow areas, priority is given to ensuring obstacle avoidance safety, thereby adapting to the needs of different operation scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the main body of the modularly designed underwater cleaning and monitoring integrated ROV method and system of the present invention, including multiple detachable modules; Figure 2 A schematic diagram showing the layout of the thrusters for the modularly designed underwater cleaning and monitoring integrated ROV method and system of the present invention. Figure 3 A schematic diagram illustrating the communication between the multi-source sensor array and the main body of the modularly designed underwater cleaning and monitoring integrated ROV method and system of this invention; Figure 4 A flowchart illustrating the modular design of the integrated underwater cleaning and monitoring ROV method and system of the present invention. Figure 5 This is a schematic diagram of the multi-scale feature refining module of the modularly designed underwater cleaning and monitoring integrated ROV method and system of the present invention; Figure 6 This is a schematic diagram of the adaptive enhancement module structure of the modularly designed integrated underwater cleaning and monitoring ROV method and system of the present invention. Figure 7This is a schematic diagram of the multi-dimensional color enhancement module of the modularly designed underwater cleaning and monitoring integrated ROV method and system of the present invention; Figure 8 The model structure diagram of the SC-YOLO algorithm for the modularly designed underwater cleaning and monitoring integrated ROV method and system of this invention is shown. Figure 9 This is a schematic diagram of the path planning of the main body of the modularly designed underwater cleaning and monitoring integrated ROV method and system of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Existing ROVs have many shortcomings: such as difficulty in disassembly and maintenance, poor flexibility in functional expansion; insufficient rationality in power layout distribution, with some ROV thruster layout design defects leading to poor motion flexibility and inability to achieve all-round precise attitude control, affecting operational efficiency; underwater imaging technology is difficult to adapt to complex aquatic environments, with prominent problems of image blurring and color distortion; target detection algorithms have insufficient accuracy in identifying low-resolution, multi-scale underwater debris targets, with a high false negative rate, resulting in the cleaning of underwater debris, such as microplastics and discarded fishing nets, still mainly relying on manual retrieval, which has the disadvantages of low efficiency and high safety hazards.

[0020] In view of this, such as Figure 1 and Figure 2 As shown, on one hand, the present invention provides a modularly designed integrated underwater cleaning and monitoring ROV system, comprising: The main body consists of several functional modules, each with its own independent function, and adjacent functional modules can be detachably connected. Several thrusters are installed at different positions on the main body to drive the main body to move in all directions underwater; A multi-source sensor array, integrated on the main body, is used to collect underwater environment data, obtain water parameters and images, and achieve multi-dimensional perception. The system incorporates a lightweight multi-scale image enhancement network and the SC-YOLO algorithm. The lightweight multi-scale image enhancement network is used to extract feature maps from images acquired by a multi-source sensor array to generate enhanced images. The SC-YOLO algorithm uses the enhanced images as input to identify underwater debris and feeds the results back to the system's functional modules to perform cleanup operations.

[0021] Thanks to its modular and scalable structure, the ROV body is divided into several functional modules through modular hardware and standardized interface design. It integrates a multi-source sensor array and a specific thruster layout, which allows for rapid maintenance and assembly, as well as multi-dimensional perception of the underwater environment. Furthermore, through the built-in lightweight multi-scale image enhancement network and SC-YOLO algorithm, it can identify underwater debris targets of different sizes. Combined with the scalable mechanical handle, it can perform cleaning operations, better achieving the goals of water purification and obstacle avoidance.

[0022] like Figure 1 As shown, in one embodiment, the functional modules constituting the main body include a main control module, a power output module, an environmental sensing module, a cleaning module, and an energy storage module. The main control module is communicatively connected to the power output module, the environmental sensing module, and the cleaning module. The energy storage module provides energy to the main control module, the power output module, the environmental sensing module, the cleaning module, and the energy storage module. Each functional module has a cuboid structure, and each side surface is provided with a slider and a groove. Adjacent functional modules directly transmit energy or communicate via aviation connectors. A multi-source sensor array is communicatively connected to the environmental sensing module. Figure 3 As shown, the multi-source sensor array includes, but is not limited to, cameras, multi-beam sonar, gyroscopes, depth sensors, and turbidity sensors. The number of sensors can be increased or decreased according to actual needs. Adjacent modules can be connected via preset sliders and grooves, and then fastened with bolts. Compared to magnetic connections, this significantly improves resistance to water flow impact, and maintenance personnel can replace modules or troubleshoot faults single-handedly. A unified electrical interface is a standard CAN bus, ensuring plug-and-play functionality. This interchangeable modular structure means that only the corresponding module needs to be replaced in case of a fault, greatly reducing maintenance and repair time. A dedicated temporary storage module can be configured to store recovered underwater debris. The cleaning module can be configured as one or more of a robotic arm, a suction pump, or a cutter to handle different types of underwater debris. For example, cutting and gripping strategies can be used for fishing nets, while adsorption and filtration can be used for microplastics, avoiding damage to underwater facilities or secondary pollution caused by broken debris, providing reliable cleaning efficiency.

[0023] like Figure 2As shown, the power output modules are located at various vertices of the main control module and on one side surface of the main control module. Each power output module is equipped with a corresponding thruster. The thrusters at the vertices of the main control module drive the ROV's planar motion underwater, while the thrusters on the side surface of the main control module adjust the ROV's depth underwater. This embodiment uses a thruster combination layout of 2 front, 2 rear, and 2 side thrusters. The thrusters at the vertices of the main control module maintain the ROV's planar motion, while the two side thrusters control the buoyancy of the ROV. In particular, the two side thrusters, in conjunction with a conventional vector deflection mechanism, can achieve omnidirectional motion with ±90° roll and ±30° lateral yaw.

[0024] On the other hand, the present invention provides a modularly designed integrated underwater cleaning and monitoring ROV method, comprising the following steps: S1: The above-mentioned modular underwater cleaning and monitoring integrated ROV system is equipped with a multi-source sensor array including a camera, multi-beam sonar, gyroscope, depth sensor and turbidity sensor, which respectively acquire underwater images, underwater obstacles, body attitude, current depth and water turbidity and feed them back to the body.

[0025] This approach employs a combination of multiple sensors to compensate for the limitations of a single sensor. For example, in turbid water, sonar can fill detection blind spots and improve the ability to detect underwater targets. To ensure the reliability of underwater operations, each sensor can be used in a one-for-one backup configuration. If the currently used sensor malfunctions or its error exceeds a set threshold, the backup sensor can be activated immediately without interrupting operations for maintenance or repair. To ensure reliable data transmission, each sensor can use I2C and SPI buses to acquire data in parallel, making the acquired data more reliable.

[0026] S2: The lightweight multi-scale image enhancement network built into the ontology performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images to enhance input characteristics and improve the overall image effect. Then, the adaptive enhancement module performs multi-scale feature modulation to achieve adaptive feature enhancement. Finally, the multi-dimensional color enhancement module enhances color information from three dimensions: channel, space, and depth, to improve color fidelity and detail sharpness, and outputs the enhanced feature map. In step S2, the lightweight multi-scale image enhancement network built into the ontology performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images to enhance input characteristics and improve the overall image quality. This is achieved by using a multi-scale feature refinement module, which includes a head section, depthwise separable convolutional units, multi-branch dilated depthwise convolutional units, feature fusion units, and a tail section. The head section includes sequentially arranged 1×1 convolutional layers and batch normalization (BN) layers. The 1×1 convolutional layers are used for dimensionality reduction to adjust the original number of channels, and the BN layers stabilize the training process by introducing non-linearity through the GELU activation function. The depthwise separable convolutional unit Dconv5×5 uses a 5×5 convolutional kernel. Each channel of the features output from the head section is convolved to process spatial information, but there is no mixing between channels. The features processed by the depthwise separable convolutional unit are then fed into a multi-branch dilated depth convolutional unit. The multi-branch dilated depth convolutional unit includes several parallel branches, each branch including pairs of dilated convolutional layers. The convolutional kernels of the dilated convolutional layers in each branch have the same depth, but different receptive fields. The feature fusion unit is used to fuse the feature maps output from several branches, so that the fused feature map contains information at different scales simultaneously. The tail section has the same structure as the head section, including 1×1 convolutional layers, batch normalization (BN) layers, and the GELU activation function, which are used to map the fused feature map back to the original number of channels, completing multi-scale feature refinement.

[0027] Traditional Unet encoders extract features solely through convolution and pooling, lacking the ability to capture multi-scale details, resulting in the loss of information such as debris edges and fine textures in underwater images. This embodiment inserts a multi-scale feature refinement module after each convolutional layer of the Unet encoder. For example... Figure 5 As shown, the multi-scale feature refinement module, from bottom to top, includes a head, depthwise separable convolutional units, multi-branch dilated deep convolutional units, feature fusion units, and a tail. This module fuses shallow detail features with deep structural features in a multi-dimensional manner, enhancing the model's ability to extract and integrate detailed and global structural information. The enhanced multi-scale modulation features are then upsampled and embedded layer by layer into the decoder of the base network, thereby generating higher-quality enhanced images. This improves the model's ability to capture key visual information in underwater images and enhances input characteristics to improve the overall image enhancement effect. Figure 5 As shown in the figure, there are three parallel feature extraction paths. The three parallel branches use different depth convolution kernels and pairs of dilated convolution layers, but have different dilation rates: d.1x1, d.7x1, d.1x11, d.11,1, d.1x21, and d.21x1.

[0028] The results of the multi-scale feature refinement module are inserted into each convolutional layer of the encoder in the UNet network. Features are then extracted through three parallel paths. The first path uses 1×1 convolution and dilated convolution to capture image details. The second path uses 3×3 convolution to extract the feature contours of the target. The third path uses 5×5 convolution and global average pooling layer to capture the overall structure of the image and environmental background information. The outputs of the three parallel paths are then weighted and summed for fusion. The weights are dynamically allocated by the attention mechanism. The output is a unified feature map FIN containing shallow detail features.

[0029] In one embodiment, this embodiment also includes a lightweight modification of the UNet network. The lightweight modification removes redundant convolutional channels in the encoder by pruning, and INT8 quantization compresses the parameter precision from 32 bits to 8 bits, reducing the number of parameters to 2.4M and increasing the inference speed to 25fps, which meets the real-time processing requirements of the ROV embedded platform.

[0030] Furthermore, the decoder of the traditional Unet network, through simple upsampling and splicing of encoder features, fails to fully consider the adaptability of features at different scales, resulting in color distortion and blurred edges in the enhanced image. Therefore, as... Figure 6 As shown, in step S2, the feature fusion module of the decoder is replaced with an adaptive enhancement module. The adaptive enhancement module performs multi-scale feature modulation to achieve adaptive feature enhancement.

[0031] The adaptive enhancement module includes a main path and a gated branch. The main path consists of a first Conv layer, a first adder, a CMBA module, a second adder, and a second Conv layer arranged sequentially. The first Conv layer performs a 1×1 convolution on the input fin for channel number reduction and dimensionality adjustment. The first adder adds the dimensionality-reduced output of the first Conv layer element-wise to the input fin, and then feeds the output into the CMBA module. The CMBA module is an attention module that receives the features input from the first adder and outputs an enhanced feature that has been recalibrated for both channel and spatial dimensions. The second adder adds the output of the CMBA module element-wise to the input fin again. The second Conv layer performs a 1×1 convolution on the features received from the second adder. Convolution maps features back to the same number of channels as the input fin. The gated branch includes sequentially arranged BN layers, PreLU layers, Pooling layers, first Linear layers, ReLU layers, second Linear layers, Sigmoid layers, and multipliers. The BN layer is used for batch normalization of the input features. The PreLU layer is an activation function with learning parameters used to introduce non-linearity. The Pooling layer is used for global average pooling, compressing the entire feature map into a scalar. The first Linear layer is used to perform dimensionality reduction and non-linear transformation on the pooled channel vectors. The ReLU layer is used to introduce non-linearity after the first Linear layer. The second Linear layer is used to restore the feature dimension to the original number of channels. The Sigmoid layer is used to map the output of the second Linear layer to the (0,1) interval to generate an adaptive weight vector δ. The multiplier multiplies the output of the CMBA module with the adaptive weight vector δ generated by the gated branch element-wise to obtain the enhanced deep structural features, which are used as the output of the adaptive enhancement module. The decoder uses dynamic convolution kernel adjustment to ensure scale matching between the encoder's downsampled features and the decoder's upsampled features. Different convolution kernels are used in the encoder, and the scale of the corresponding upsampled feature map in the decoder adaptively switches to the appropriate kernel size, ensuring precise scale matching between the encoder's downsampled features and the decoder's upsampled features, and avoiding blurring caused by feature misalignment. The gating branch dynamically calibrates the weights of the shallow detail features transmitted by the encoder and the deep structural features generated by the decoder. Dynamic parameter calibration combines the turbidity values ​​from the turbidity sensor to dynamically adjust the weight allocation strategy for the shallow detail features transmitted by the encoder and the deep structural features generated by the decoder. Specifically: , NTU The value is a unit of turbidity. In high turbidity environments, shallow details have a higher weight, which can enhance edge sharpening; in low turbidity environments, deep structural features have a higher weight, which can ensure natural image colors.

[0032] like Figure 7As shown, in a preferred embodiment, due to the special nature of underwater images, the deeper the sea area, the more difficult it is to capture images, and the more easily colors are distorted. To improve the efficiency of capturing and processing visual information about the underwater environment, step S2 further enhances color information from three dimensions: channel, space, and depth using a multi-dimensional color enhancement module. This improves color fidelity and detail sharpness, outputting an enhanced feature map. Specifically, the multi-dimensional color enhancement module adaptively enhances features in the three dimensions of color, channel, and space. The multi-dimensional color enhancement module includes a multi-scale feature fusion stage, a multi-dimensional attention feature enhancement stage, and a feature reconstruction and residual output stage. The multi-scale feature fusion stage includes a first BatchNorm layer, several parallel convolutional layers, a first channel stitching layer C1, and a feature transformation sequence step. The first BatchNorm layer is used for batch normalization, refining the multi-scale features. The result of the block is used as the current input feature for standardization. Several parallel convolutional layers are used to extract features of different scales from the standardized current input features in parallel. Through parallel 1×1 convolution, 5×5 convolution and 7×7 convolution, features of different scales in the image are perceived simultaneously. The first channel stitching layer C1 is used to stitch the feature maps output by several parallel convolutional layers in the channel dimension to form a multi-scale feature tensor. The feature transformation sequence includes a first 1×1 Conv layer, a first GELU layer, a second 1×1 Conv layer and a third adder set in sequence, which are used to reduce the number of channels, introduce nonlinearity and map the features back to the original number of channels, respectively, and then add them to the current input features element by element.

[0033] The multi-dimensional attention feature enhancement stage includes parallel Channel-wise Attention, Spatial-wise Attention, 3D-wise Attention, and a second channel stitching layer C2. Channel-wise Attention dynamically adjusts the gain of each RGB channel based on its attenuation characteristics. Spatial-wise Attention performs differential color correction on the target and background regions of the image. 3D-wise Attention combines depth sensor data to adjust the color compensation coefficient based on the current depth of the image. The second channel stitching layer C2 weights and stitches the outputs of Channel-wise Attention, Spatial-wise Attention, and 3D-wise Attention in the channel dimension, aggregating enhanced features from the channel, spatial, and joint 3D perspectives. The multi-dimensional attention feature enhancement stage also constructs a dynamic water parameter adaptation mechanism to dynamically adjust the weights of each branch of Channel-wise Attention, Spatial-wise Attention, and 3D-wise Attention. The feature reconstruction and residual output stage includes a second BatchNorm layer, a third 1×1 Conv layer, a second GELU layer, a fourth 1×1 Conv layer, and a fourth adder. The second BatchNorm layer is located between the multi-scale feature fusion stage and the multi-dimensional attention feature enhancement stage, and is used for batch normalization again. The third 1×1 Conv layer, the second GELU layer, the fourth 1×1 Conv layer, and the fourth adder are located after the multi-dimensional attention feature enhancement stage. They are used to reduce the dimensionality of the result after the second channel splicing layer C2, introduce nonlinearity again, and perform the final projection to map the features back to the same dimension as the current input features. The fourth adder then adds the feature element-wise to the current input features.

[0034] The dynamic adaptation mechanism for water parameters involves using turbidity and depth sensors to collect data on water turbidity and the depth of the operating system, and then determining the following branch weights: 1) Turbidity > 300 NTUFor depths > 30 meters, the weight of Channel-wise Attention is increased to 60%, and the combined weights of Spatial-wise Attention and 3D-wise Attention are increased to 40%. 2) 150 NTU ≤Turbidity≤300 NTU For depths of 10 meters to 30 meters, the weight of Channel-wise Attention is 35%, the weight of Spatial-wise Attention is 35%, and the weight of 3D-wise Attention is 30%. 3) Turbidity <150 NTU For depths <10 meters, the weight of the Spatial-wise Attention branch is 55%; the sum of the weights of the Channel-wise Attention branch and the 3D-wise Attention branch is 50%.

[0035] The multi-dimensional color enhancement module introduces a depth dimension and constructs a three-dimensional enhancement architecture of channel-space-depth. In the channel dimension, it dynamically adjusts the gain of each channel based on the attenuation characteristics of the RGB three channels, with red light attenuating the fastest and blue light attenuating the slowest. In the spatial dimension, it performs differentiated color correction on local areas of the image, such as the target area and the background area, to avoid the reduction in target and background contrast caused by global correction. In the depth dimension, it combines depth sensor data and adjusts the color compensation coefficient according to the ROV operation depth.

[0036] The dynamic water parameter adaptation mechanism establishes a mapping relationship between turbidity, depth, and color compensation coefficients in advance through training with experimental data. It then collects real-time data on water turbidity and the operating depth of the instrument, dynamically adjusting branch weights accordingly. This allows for adjustments in turbidity ranges from 50 to 500. NTU Image clarity can be maintained within the range, and the effective observation distance is extended from 2 meters in the traditional method to 5 meters.

[0037] S3: Use the improved SC-YOLO algorithm architecture to perform underwater debris target detection and output target localization results.

[0038] like Figure 8As shown, step S3 involves replacing the C2f module in the feature fusion network of the YOLOv8 algorithm architecture with the C2f_SCConv module. The C2f module is known for its powerful feature extraction and fusion capabilities, while the SCConv module excels in feature representation and detail capture. Therefore, the C2f_SCConv module not only inherits the advantages of both modules but also significantly improves the model's detection performance and feature representation capabilities through complementary effects. This improvement brings significant advantages to YOLOv8, especially when dealing with complex and varied detection tasks. It effectively improves the model's detection accuracy and robustness in these complex scenarios. Simultaneously, this structure further enhances the feature representation capability, enabling the model to more accurately utilize feature information in subsequent detection and recognition tasks, thereby improving the overall task execution efficiency and accuracy.

[0039] The core innovation of the C2f_SCConv module lies in its deep coupling of the efficient feature fusion capability of the C2f module with the fine-grained feature calibration capability of the self-calibrating convolutional SCConv module. This solves the problems of traditional modules in extracting features in complex underwater environments, where the features of small targets are easily submerged. Unlike conventional improvements that simply replace modules, this module achieves an integrated process of feature extraction, calibration, and fusion through structured reconstruction, specifically adapted to scenarios where details remain blurred and the contrast between the target and the background is low even after underwater image enhancement.

[0040] The C2f_SCConv module specifically includes an input splitting unit, a C2f basic feature extraction unit, an SCConv self-calibration feature optimization unit, and a feature fusion and aggregation unit. The input splitting unit further splits the input feature map into two parallel main feature branches and calibration feature branches after passing it through a continuously set Conv-batch normalized BN-SiLU activation function. The main feature branch is used for global feature extraction of the feature map, and the calibration feature branch is used for local detail calibration of the feature map. The ratio of the main feature branch to the calibration feature branch is 1:1, ensuring a balanced capture of global features and local detail features. The C2F basic feature extraction unit contains sequentially arranged Bottleneck sub-units. Each Bottleneck sub-unit includes 1×1 convolutional dimensionality reduction, 3×3 convolutional feature extraction, and residual connection structure, which is suitable for underwater garbage targets with a size of 5cm-50cm. The SCConv self-calibration feature optimization unit specifically includes an SRU spatial calibration submodule and a CRU channel calibration submodule. The SRU spatial calibration submodule uses a dynamic receptive field adjustment mechanism to perform multi-scale window sliding sampling on the feature map, capturing fine-grained features such as target edges and textures, thus improving the problem of indistinct features of small underwater targets, such as microplastics. The CRU channel calibration submodule uses an attention mechanism to dynamically allocate feature channel weights, generates channel attention maps through the Sigmoid activation function, strengthens the features of channels related to debris targets, and suppresses background noise channels. The feature fusion and aggregation unit fuses the features of the main feature branch and the calibration feature branch through element-wise addition and pointwise convolution, then unifies the feature dimensions through a 1×1 convolutional layer, and finally enhances the non-linear expressive power of the features through the SiLU activation function, outputting the calibrated feature map.

[0041] Taking underwater plastic waste as an example, the specific working process of the C2f_SCConv module is as follows: The underwater image feature map after multi-scale image enhancement, with a size of 640×640×256, is input into the C2F_SCConv module. It is split into two branches through the CBS layer, and the output feature map size of each branch is maintained at 640×640×128. The main feature branch extracts the global contour features of microplastics step by step through the Bottleneck stacking structure of C2f, while retaining shallow detail information through residual connections to avoid feature gradient vanishing, and outputs a feature map containing the global structure. In the calibration feature branch, the spatial calibration submodule captures the highlight texture of the microplastic surface through 3×3 window sampling and enhances the boundary distinction between microplastics and the surrounding water through 5×5 window sampling. The channel calibration submodule increases the weight of the corresponding channel for the RGB color features of microplastics, suppresses the interference of the blue-green background channel of the water, and outputs a precisely calibrated local feature map. The feature maps of the two branches are element-wise added to achieve feature complementarity. The feature dimension is compressed to 640×640×256 through 1×1 convolution. The SiLU activation function enhances the non-linear expression of the features. The final output is a highly recognizable feature map that contains both global contours and local details, providing accurate feature input for subsequent detection heads.

[0042] S4: Based on the positioning results and the body attitude obtained by the gyroscope, the dynamic path planning model improved by the A* algorithm is used to plan the path of the body, so as to realize the obstacle avoidance path planning of the body underwater.

[0043] In this embodiment, to improve the accuracy and robustness of path planning, the target localization results output by the SC-YOLO algorithm and the body attitude obtained by the gyroscope are integrated, and a dynamic path planning model improved by the A* algorithm is used for path planning. In areas with dense waste, priority is given to approaching the target; in narrow areas, priority is given to ensuring obstacle avoidance safety, adapting to the needs of different operational scenarios.

[0044] The improved dynamic path planning model based on the A* algorithm includes the following: 1) Introducing a dynamic cost function to dynamically adjust path weights according to the water environment. For example, when there is a sudden change in illumination, the stability of inertial sensor data is increased in the cost calculation to compensate for the decrease in the reliability of visual data; 2) Adopting a hierarchical search mechanism to quickly generate a general global path first, and then to refine the local path based on real-time sensor data, balancing planning efficiency and accuracy; 3) Using a distance-velocity dual-threshold obstacle avoidance logic, which monitors the distance between obstacles and the ROV body and the obstacle's moving speed in real time through multi-source sensors. When the distance is lower than the safety threshold or the obstacle's moving speed is higher than the set threshold, emergency obstacle avoidance measures are taken, and the path is recalculated after the emergency obstacle avoidance measures are triggered; 4) Prioritizing approaching the target in areas with dense garbage and ensuring obstacle avoidance safety in narrow areas to adapt to the needs of different operating scenarios.

[0045] In this embodiment, the dynamic cost function is: , From the starting point to the current node n The estimated cost, Weighting terms , The cost of visual estimation is inversely proportional to the image sharpness. The cost to the stability of the body's attitude is related to the stability of the inertial sensor data. This represents the prior cost based on static maps; It is the estimated cost from the current node to the destination. The penalty coefficient is... For the current node n Distance to the nearest obstacle It is a non-zero constant.

[0046] S5: Upon reaching the underwater debris site, the main unit drives the cleaning module to clean up the underwater debris.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A modular design for an integrated underwater cleaning and monitoring ROV method, characterized in that, Includes the following steps: S1: An integrated ROV system for underwater cleaning and monitoring with a modular design, including the main body and a multi-source sensor array; the multi-source sensor array specifically includes a camera, a multi-beam sonar, a gyroscope, a depth sensor and a turbidity sensor, which respectively acquire underwater images, underwater obstacles, the attitude of the main body, the current depth and the turbidity of the water and feed them back to the main body; S2: The lightweight multi-scale image enhancement network built into the ontology performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images to enhance input characteristics and improve the overall image effect. Then, the adaptive enhancement module performs multi-scale feature modulation to achieve adaptive feature enhancement. Then, the multi-dimensional color enhancement module enhances color information from three dimensions: channel, space, and depth, improving color fidelity and detail sharpness, and outputting an enhanced feature map; S3: Use the improved SC-YOLO algorithm architecture to perform underwater debris target detection and output the target localization results; S4: Based on the positioning results and the body attitude obtained by the gyroscope, the dynamic path planning model improved by the A* algorithm is used to plan the path of the body, so as to realize the obstacle avoidance path planning of the body underwater. S5: Upon reaching the underwater debris site, the main unit drives the cleaning module to clean up the underwater debris.

2. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 1, characterized in that, The lightweight multi-scale image enhancement network built into the ontology described in step S2 performs multi-dimensional fusion of shallow detail features and deep structural features of underwater images, enhancing input characteristics and improving the overall image quality. This is achieved by using a multi-scale feature refinement module to process the image. This module includes a head, depthwise separable convolutional units, multi-branch dilated depthwise convolutional units, feature fusion units, and a tail, arranged sequentially. The head includes sequentially arranged 1×1 convolutional layers and batch normalization (BN) layers. The 1×1 convolutional layers are used for dimensionality reduction to adjust the original number of channels. The normalized layer BN stabilizes the training process by introducing nonlinearity through the GELU activation function. The depthwise separable convolutional unit Dconv5×5 uses a 5×5 convolutional kernel to perform convolution operations on each channel of the features output from the head, processing spatial information, but without mixing between channels. The features processed by the depthwise separable convolutional unit are then fed into a multi-branch dilated deep convolutional unit. The multi-branch dilated deep convolutional unit includes several parallel branches, each branch including pairs of dilated convolutional layers. The convolutional kernels of the dilated convolutional layers in each branch have the same depth, but different receptive fields. The feature fusion unit is used to fuse the feature maps output from several branches, so that the fused feature map contains information at different scales simultaneously. The tail has the same structure as the head, including a 1×1 convolutional layer, a batch normalization layer (BN), and a GELU activation function, which are used to map the fused feature map back to the original number of channels, thus completing multi-scale feature refinement.

3. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 2, characterized in that, The step S2, which involves performing multi-scale feature modulation through an adaptive enhancement module to achieve adaptive feature enhancement, is the result of the multi-scale feature refinement module. After inserting it into each convolutional layer of the encoder in the UNet network, features are extracted through three parallel paths. The first path uses 1×1 convolution and dilated convolution to capture image details, the second path uses 3×3 convolution to extract the feature contours of the target, and the third path uses 5×5 convolution and a global average pooling layer to capture the overall structure of the image and environmental background information. The outputs of the three parallel paths are then weighted and summed for fusion. The weights are dynamically allocated by the attention mechanism, and the output is an integrated feature map containing shallow detail features. Then, at the decoder end of the UNet network, the original feature fusion module in the decoder structure is replaced with an adaptive enhancement module. The adaptive enhancement module includes a main path and a gated branch. The main path includes a first Conv layer, a first adder, a CMBA module, a second adder, and a second Conv layer arranged sequentially. The first Conv layer performs a 1×1 convolution on the input fin for channel number reduction adjustment. The first adder adds the dimensionality-reduced output of the first Conv layer to the input fin element-wise, and then sends the output to the CMBA module. The CMBA module is an attention module that receives the features input from the first adder and outputs an enhanced feature that has been recalibrated for both channel and spatial dimensions. The second adder adds the output of the CMBA module to the input fin again element-wise. The second Conv layer performs a 1×1 convolution on the features received from the second adder output, mapping the features back to the same number of channels as the input fin; the gated branch includes a BN layer, a PreLU layer, a Pooling layer, a first Linear layer, a ReLU layer, a second Linear layer, a Sigmoid layer, and a multiplier arranged in sequence. The BN layer is used to perform batch normalization on the input features. A PreLU layer is an activation function with learned parameters used to introduce nonlinearity; The pooling layer is used for global average pooling, compressing the entire spatial feature map into a scalar; the first linear layer is used to perform dimensionality reduction and nonlinear transformation on the pooled channel vectors; and the ReLU layer is used to introduce nonlinearity after the first linear layer. The second Linear layer is used to restore the feature dimension to the original number of channels; the Sigmoid layer is used to map the output of the second Linear layer to the (0,1) interval to generate an adaptive weight vector δ; the output of the CMBA module is multiplied element-wise with the adaptive weight vector δ generated by the gated branch through a multiplier to obtain the enhanced deep structural features, which are used as the output of the adaptive enhancement module. The decoder uses dynamic convolution kernel adjustments to ensure scale matching between the encoder's downsampled features and the decoder's upsampled features.

4. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 3, characterized in that, This also includes lightweight modifications to the UNet network and dynamic parameter calibration. Lightweight modifications involve removing redundant convolutional channels from the encoder; dynamic parameter calibration involves dynamically adjusting the weight allocation strategy between the shallow detail features transmitted by the encoder and the deep structural features generated by the decoder, based on the turbidity values ​​from the turbidity sensor. The specific details are as follows: , NTU It is a unit of turbidity.

5. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 3, characterized in that, The step S2 described above involves using a multi-dimensional color enhancement module to enhance color information from three dimensions: channel, space, and depth, thereby improving color fidelity and detail sharpness, and outputting an enhanced feature map. This step involves using a multi-dimensional color enhancement module to adaptively enhance features in the three dimensions of color, channel, and space. The multi-dimensional color enhancement module includes a multi-scale feature fusion stage, a multi-dimensional attention feature enhancement stage, and a feature reconstruction and residual output stage; The multi-scale feature fusion stage includes a first BatchNorm layer, several parallel convolutional layers, a first channel concatenation layer C1, and a feature transformation sequence. The first BatchNorm layer is used for batch normalization, standardizing the results of the multi-scale feature refinement module as the current input features. Several parallel convolutional layers are used to extract features of different scales from the standardized current input features in parallel. Through parallel 1×1 convolutions, 3×3 convolutions, and 5×5 convolutions, features of different scales in the image are perceived simultaneously. The first channel concatenation layer C1 is used to concatenate the feature maps output by several parallel convolutional layers along the channel dimension to form a multi-scale feature tensor. The feature transformation sequence includes a first 1×1 Conv layer, a first GELU layer, a second 1×1 Conv layer, and a third adder, which are set sequentially to reduce the number of channels, introduce nonlinearity, and map the features back to the original number of channels, respectively, before being added element-wise with the current input features. The multi-dimensional attention feature enhancement stage includes parallel channel-wise attention, spatial-wise attention, 3D-wise attention, and a second channel stitching layer C2. The channel-wise attention dynamically adjusts the gain of each RGB channel based on the attenuation characteristics of the three RGB channels. The spatial-wise attention performs differential color correction on the target and background regions in the image. The 3D-wise attention combines depth sensor data and adjusts the color compensation coefficient according to the current working depth of the ontology. The second channel stitching layer C2 weights and stitches the outputs of the parallel channel-wise attention, spatial-wise attention, and 3D-wise attention in the channel dimension, aggregating the enhanced features from the channel, spatial, and joint 3D perspectives. In the multidimensional attention feature enhancement stage, a dynamic water parameter adaptation mechanism is also constructed to dynamically adjust the weights of each branch of the channel-wise attention branch, the spatial-wise attention branch, and the three-dimensional attention branch, 3D-wise attention. The feature reconstruction and residual output stage includes a second BatchNorm layer, a third 1×1 Conv layer, a second GELU layer, a fourth 1×1 Conv layer, and a fourth adder. The second BatchNorm layer is located between the multi-scale feature fusion stage and the multi-dimensional attention feature enhancement stage, and is used for batch normalization again. The third 1×1 Conv layer, the second GELU layer, the fourth 1×1 Conv layer, and the fourth adder are located after the multi-dimensional attention feature enhancement stage. They are used to reduce the dimensionality of the result after the second channel splicing layer C2, introduce nonlinearity again, and perform the final projection to map the features back to the same dimension as the current input features. The fourth adder then adds the feature element-wise to the current input features.

6. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 5, characterized in that, The dynamic adaptation mechanism for water parameters uses turbidity sensors and depth sensors to collect data on water turbidity and the operating depth of the system, and determines the following branch weights: 1) Turbidity > 300 NTU For depths > 30 meters, the weight of Channel-wise Attention is increased to 60%, and the combined weights of Spatial-wise Attention and 3D-wise Attention are increased to 40%. 2) 150 NTU ≤Turbidity≤300 NTU For depths of 10 meters to 30 meters, the weight of Channel-wise Attention is 35%, the weight of Spatial-wise Attention is 35%, and the weight of 3D-wise Attention is 30%. 3) Turbidity <150 NTU For depths <10 meters, the weight of the Spatial-wise Attention branch is 55%; the sum of the weights of the Channel-wise Attention branch and the 3D-wise Attention branch is 50%.

7. The modular design of the integrated underwater cleaning and monitoring ROV method according to claim 5, characterized in that, Step S3 is as follows: The SC-YOLO algorithm replaces the C2f module in the feature fusion network of the YOLOv8 algorithm architecture with the C2f_SCConv module. The C2f_SCConv module specifically includes an input splitting unit, a C2F basic feature extraction unit, an SCConv self-calibration feature optimization unit, and a feature fusion and aggregation unit. The input splitting unit is used to further split the input feature map into two parallel main feature branches and calibration feature branches after passing it through a continuously set convolutional Conv-batch normalized BN-SiLU activation function. The main feature branch is used to extract global features from the feature map, and the calibration feature branch is used to calibrate local details of the feature map. The ratio of the main feature branch to the calibration feature branch is 1:

1. The C2F basic feature extraction unit contains sequentially arranged Bottleneck sub-units. Each Bottleneck sub-unit includes 1×1 convolutional dimensionality reduction, 3×3 convolutional feature extraction, and residual connection structure, which is suitable for underwater garbage targets with a size of 5cm-50cm. The SCConv self-calibrated feature optimization unit performs multi-scale window sliding sampling on the feature map through dynamic field perception adjustment, capturing the edge and texture features of underwater debris targets and suppressing background noise; The feature fusion and aggregation unit fuses the features of the main feature branch and the calibration feature branch through element-wise addition and pointwise convolution, then unifies the feature dimensions through a 1×1 convolutional layer, and finally enhances the non-linear expressive power of the features through the SiLU activation function, outputting the calibrated feature map.

8. A modularly designed integrated underwater cleaning and monitoring ROV system, characterized in that, include: The main body consists of several functional modules, each with its own independent function, and adjacent functional modules can be detachably connected. Several thrusters are installed at different positions on the main body to drive the main body to move in all directions underwater; A multi-source sensor array, integrated on the main body, is used to collect underwater environment data, obtain water parameters and images, and achieve multi-dimensional perception. The system incorporates a modularly designed integrated underwater cleaning and monitoring ROV method as described in any one of claims 1-7. A lightweight multi-scale image enhancement network is used to extract feature maps from images acquired by a multi-source sensor array to generate enhanced images. The SC-YOLO algorithm uses the enhanced images as input to identify underwater debris and feeds it back to the functional modules of the system to perform cleaning operations.

9. The modularly designed integrated underwater cleaning and monitoring ROV system according to claim 8, characterized in that, Several functional modules include a main control module, a power output module, an environmental sensing module, a cleaning module, and an energy storage module. The main control module is communicatively connected to the power output module, the environmental sensing module, and the cleaning module. The energy storage module provides energy to the main control module, the power output module, the environmental sensing module, the cleaning module, and the energy storage module. Each functional module has a cuboid structure, with sliders and grooves on each side surface. Adjacent functional modules directly transmit energy or communicate via aviation connectors. A multi-source sensor array is communicatively connected to the environmental sensing module.

10. The modularly designed integrated underwater cleaning and monitoring ROV system according to claim 9, characterized in that, The power output modules are located at each vertex position of the main control module and on one side surface of the main control module. Each power output module is equipped with a corresponding thruster. The thrusters at each vertex position of the main control module are used to drive the body's planar motion underwater, while the thrusters on the side surface of the main control module are used to adjust the body's depth underwater.

Images