Submarine organism target detection method based on RFM-DETR

Abstract

The invention discloses a seabed organism target detection method based on RFM-DETR, and relates to the technical field of target detection. Comprising the following steps: acquiring a benthic organism image data set, performing enhancement processing on the benthic organism image data set, and inputting the benthic organism image data set into an RFM-DETR target detection algorithm for training to obtain a benthic organism target detection model; and inputting the to-be-identified benthic organism image into the benthic organism target detection model to obtain a to-be-identified benthic organism target detection result. According to the method, the data set is acquired and enhanced, so that denoising and definition restoration of the benthic organism image are realized, and the image quality is improved.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically to a method for detecting marine biological targets based on RFM-DETR. Background Technology

[0002] In complex marine environments, underwater images are often blurry, have low contrast, and are subject to background interference from seaweed, plankton, and other organisms, making feature extraction difficult and resulting in high false negative rates for deep learning-based target detection algorithms. The small size and low contrast of small and medium-sized marine organisms further complicate detection. Furthermore, mobile devices are limited by computing power, memory, and battery life, requiring detection algorithms with excellent performance while balancing model complexity and computational efficiency. Therefore, developing efficient, lightweight, and robust target detection algorithms is crucial for advancing underwater detection technology.

[0003] In the field of marine exploration, underwater target detection technology is particularly crucial, with a wide range of applications including environmental monitoring, marine ecological research, underwater archaeology, and military reconnaissance. Among these, the study of deep-sea fish is not only related to ecosystem balance, but their population size and distribution patterns also have significant indicative value for environmental monitoring and climate change research. However, in current scientific exploration, the accurate detection and identification of deep-sea fish still faces many severe challenges. The lighting conditions in the deep-sea environment are extremely complex and variable, easily affected by the interplay of multiple factors such as seasonal changes, weather conditions, and ocean current dynamics. This directly leads to the instability of the appearance characteristics of target fish, thus significantly increasing the technical difficulty of target detection.

[0004] To address the shortcomings in feature extraction and fusion in the detection of small marine organisms, as well as the problems of false detection and missed detection of small targets caused by image blur and complex backgrounds, this paper proposes a marine organism target detection method based on RFM-DETR to overcome the difficulties of existing technologies. This is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method for detecting marine biological targets based on RFM-DETR, which is used to solve the technical problems existing in the prior art.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting marine biological targets based on RFM-DETR includes the following steps:

[0008] S1. Obtain a dataset of marine life images and perform enhancement processing on the dataset;

[0009] S2. Input the seabed organism image dataset into the RFM-DETR target detection algorithm for training to obtain the seabed organism target detection model;

[0010] S3. Input the image of the seabed organism to be identified into the seabed organism target detection model to obtain the detection result of the seabed organism target to be identified.

[0011] Optionally, the specific content of the enhancement processing of the seabed organism image dataset in S1 is as follows:

[0012] An image enhancement method based on the Transformer diffusion model is used to enhance a dataset of marine life images. The diffusion model includes a forward process and a backward process. The forward process is a Markov process that gradually adds normal Gaussian noise to destroy the image, gradually removing image details until it becomes pure noise. In the backward process, pure noise is input and a neural network is trained to gradually remove the noise introduced by the forward process.

[0013] Optionally, the specific details of the reverse process are as follows:

[0014] Introducing the conditional diffusion model, in which, given input (x) t (,c,t), where x t The image is a noisy image, c represents the conditional image, t represents the time step, and the denoising network is used to estimate the noise distribution ∈ t During training, the L loss function is used to optimize the denoising network, as shown in the following formula:

[0015] L s =||∈ t -∈ θ (x t ,c,t)||

[0016] Where, ∈ θ The image is a predicted noisy image. The inverse process in the diffusion model is an iterative denoising process or an inverse Markov process, i.e., estimating q(x). t-1 |x t First, use pure Gaussian noise x T The estimation is performed using the following formula:

[0017]

[0018] Where, p θ (x 0:T |c) is given the conditional image c, from the noise x T The joint probability distribution of generating a clean image x0 step by step, p(x T ) is the initial Gaussian noise x T The prior distribution of p θ (xt-1 |x t c) is at step t, given the noisy image x t Given the conditional image c, predict the previous image x. t-1 The probability distribution is assumed to be Gaussian, μ θ (x t ,c,t) is the mean, ∑ θ (x t (c,t) represents the variance, and T is the total number of time steps, i.e., the number of time steps required for the diffusion process to go from a pure noise state to a clear image state. It is a Gaussian distribution;

[0019] When β t When the size is very small, the reverse process of each step is treated as a Gaussian process, and a network is designed to estimate the mean μ. θ (·) and variance ∑ θ (·), the formula for the average value is as follows:

[0020]

[0021] By sampling a Gaussian distribution using a reparameterization technique, given a Gaussian distribution... Next, sample x using z = μ + σ⊙∈, where ∈ is a normal Gaussian distribution and ⊙ is the element-wise product; finally, iteratively generate sample images using the following formula until the final image x0 is obtained, as follows:

[0022]

[0023] Optionally, the RFM-DETR object detection algorithm includes: an RFNet backbone network, an efficient hybrid encoder based on a multi-scale convolutional attention mechanism, and an RTDETR decoder.

[0024] Optionally, the RFNet backbone network is an improved version of RFNet, with the core architecture based on ResNet50. All residual blocks are replaced with Faster blocks. Faster blocks not only have the residual structure of ResNet, but also incorporate some convolutional PCConv in each Faster block.

[0025] Optionally, the multi-scale convolutional attention mechanism includes the following specific components:

[0026] First, local information is integrated using depthwise convolution techniques;

[0027] Secondly, multi-branch deep strip convolution design is used to obtain multi-scale contextual information;

[0028] Finally, a 1×1 convolution is used to simulate and adjust the correlation between different feature channels. The output is directly used as attention weights to readjust the input features of the multi-scale convolutional attention mechanism. The formula for the multi-scale convolutional attention mechanism is as follows:

[0029]

[0030] Where F represents the input feature, and Att and Out represent the attention matrix and output, respectively. DW-Conv represents element-wise matrix multiplication, DW-Conv represents depthwise convolution, and Scale represents scale. i , i∈{0,1,2,3}, represents the i-th branch.

[0031] Optionally, two depthwise strip convolutions can be used in each branch.

[0032] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for detecting seabed biological targets based on RFM-DETR, the beneficial effects of which are:

[0033] 1) Acquire the dataset and perform enhancement processing on the dataset to achieve denoising and sharpness restoration of marine life images, thereby improving image quality;

[0034] 2) By designing the RFNet backbone network and an efficient hybrid encoder based on multi-scale convolutional attention, the algorithm performance is significantly improved. Attached Figure Description

[0035] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 A flowchart of a method for detecting marine biological targets based on RFM-DETR provided by the present invention;

[0037] Figure 2 An iterative refinement diagram of the diffusion model provided by this invention;

[0038] Figure 3 The Transformer module diagram provided for this invention is shown below; where (a) represents the denoising network and (b) represents the channel attention module.

[0039] Figure 4The underwater image enhancement effect based on the Transformer diffusion model provided by the present invention is shown in the figure; where (a) represents objects detected at close range and (b) represents objects detected at long range.

[0040] Figure 5 This is a diagram of the RFM-DETR architecture provided by the present invention;

[0041] Figure 6 This is a diagram of the RFNet architecture provided by the present invention;

[0042] Figure 7 The structure diagram of the efficient hybrid encoder based on the multi-scale convolutional attention mechanism provided by this invention is shown.

[0043] Figure 8 This is a schematic diagram of the multi-scale convolutional attention mechanism provided by the present invention;

[0044] Figure 9 A scatter plot comparing the algorithm of this invention with mainstream detectors. Detailed Implementation

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] See Figure 1 As shown, this invention discloses a method for detecting seabed biological targets based on RFM-DETR, comprising the following steps:

[0047] S1. Obtain a dataset of marine life images and perform enhancement processing on the dataset;

[0048] S2. Input the seabed organism image dataset into the RFM-DETR target detection algorithm for training to obtain the seabed organism target detection model;

[0049] S3. Input the image of the seabed organism to be identified into the seabed organism target detection model to obtain the detection result of the seabed organism target to be identified.

[0050] Furthermore, the specific details of the augmentation processing of the seabed organism image dataset in S1 are as follows:

[0051] An image enhancement method based on the Transformer diffusion model is used to enhance a dataset of marine life images. The diffusion model includes a forward process and a backward process. The forward process is a Markov process that gradually adds normal Gaussian noise to destroy the image, gradually removing image details until it becomes pure noise. In the backward process, pure noise is input and a neural network is trained to gradually remove the noise introduced by the forward process.

[0052] Furthermore, the specific details of the reverse process are as follows:

[0053] Introducing the conditional diffusion model, in which, given input (x) t (,c,t), where x t The image is a noisy image, c represents the conditional image, t represents the time step, and the denoising network is used to estimate the noise distribution ∈ t During training, the L loss function is used to optimize the denoising network, as shown in the following formula:

[0054] L s =||∈ t -∈ θ (x t ,c,t)||

[0055] Where, ∈ θ The image is a predicted noisy image. The inverse process in the diffusion model is an iterative denoising process or an inverse Markov process, i.e., estimating q(x). t-1 |x t First, use pure Gaussian noise x T The estimation is performed using the following formula:

[0056]

[0057] Where, p θ (x 0:T |c) is given the conditional image c, from the noise x T The joint probability distribution of generating a clean image x0 step by step, p(x T ) is the initial Gaussian noise x T The prior distribution of p θ (x t-1 |x t c) is at step t, given the noisy image x t Given the conditional image c, predict the previous image x. t-1 The probability distribution is assumed to be Gaussian, μ θ (x t ,c,t) is the mean, ∑ θ (x t(c,t) represents the variance, and T is the total number of time steps, i.e., the number of time steps required for the diffusion process to go from a pure noise state to a clear image state. It is a Gaussian distribution;

[0058] When β t When the size is very small, the reverse process of each step is treated as a Gaussian process, and a network is designed to estimate the mean μ. θ (·) and variance ∑ θ (·), the formula for the average value is as follows:

[0059]

[0060] By sampling a Gaussian distribution using a reparameterization technique, given a Gaussian distribution... Next, sample x using z = μ + σ⊙∈, where ∈ is a normal Gaussian distribution and ⊙ is the element-wise product; finally, iteratively generate sample images using the following formula until the final image x0 is obtained, as follows:

[0061]

[0062] Specifically, the diffusion model based on Transformer mainly includes the diffusion process (see [link]). Figure 2 and Figure 3 As shown, this is a Transformer-based denoising network and a channel attention module. Figure 3 Part (a) in the diagram is a denoising network used to encode and refine features. Figure 3 Part (b) in the diagram is the channel attention module. The Transformer-based diffusion model is based on the noisy image x. T The process involves feeding a conditional image c and a time step t to progressively generate a sharper image. The goal is to generate a corresponding enhanced image from a low-quality image. However, using the original diffusion model with uncertain results is insufficient. Therefore, a conditional diffusion model is introduced.

[0063] Specifically, underwater image enhancement using a Transformer-based diffusion model will be analyzed from both qualitative and quantitative perspectives to assess its effectiveness.

[0064] From a qualitative perspective, see Figure 4 It is evident that the augmented images, including those of sea urchins, starfish, sea cucumbers, and scallops, are significantly clearer than the original images and retain more detailed information, such as objects, textures, and features in the underwater environment. Among these, Figure 4 (a) in the diagram represents close-range detection. Figure 4(b) represents long-distance detection. From these two dimensions, the Transformer-based diffusion model also enhances the clarity of distant target objects and increases contrast, making the objects in the image more prominent and clear, which is beneficial to improving the detection accuracy of small targets.

[0065] Table 1. Quantitative analysis of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM)

[0066] Picture number PSNR SSIM 000001 22.44db 0.93 … … … 005542 20.71db 0.86 005543 18.61db 0.87 Total: 5543 Average PSNR: 24.76 dB 0.88

[0067] The quantitative analysis in Table 1 shows that the results of obtaining an average peak signal-to-noise ratio (PSNR) of 24.76 and an average structural similarity index (SSIM) of 0.88 are quite significant. A PSNR of 24.76 indicates that the enhanced image has only a small average information loss compared to the original image. The underwater image enhancement method based on the Transformer diffusion model performs excellently in preserving image quality. An average SSIM of 0.88, close to 1, indicates that the enhanced image has high similarity to the original image in terms of brightness, contrast, and structure, demonstrating good performance in maintaining image structure and detail.

[0068] During model training, factors such as light refraction, water interference, and target occlusion in the underwater environment lead to data imbalance and poor generalization. To enhance the model's robustness and better adapt to the complexity and uncertainty of the underwater environment, a Mosaic augmentation strategy is employed. This strategy randomly stitches together multiple images to increase the diversity of the dataset, especially for small targets and occluded targets. Simultaneously, to address the issue of low detection accuracy caused by target overlap and occlusion due to the clustering of underwater organisms, a data-agnostic augmentation method—Mixup—is used. This method generates new training samples by linearly combining different images and labels, effectively reducing the impact of mislabeled images on model training and helping the model generalize better to unknown data.

[0069] Furthermore, the RFM-DETR object detection algorithm includes: an RFNet backbone network, an efficient hybrid encoder based on a multi-scale convolutional attention mechanism, and an RTDETR decoder.

[0070] For details on the overall architecture of RFM-DETR, please refer to [link / reference]. Figure 5As shown, the gray modules between P5 and P3, and P3 and P4, represent multi-scale convolutional attention mechanisms. This invention designs RFNet as the backbone network, achieving a balance between feature extraction, speed, and computational efficiency. Furthermore, an efficient hybrid encoder based on multi-scale convolutional attention mechanisms is designed. Multi-scale convolutional attention mechanisms are added to the P5 and P4 feature maps of the efficient hybrid encoder to more effectively encode contextual information, thereby achieving information aggregation from low to high levels. For the detection head, the decoder design of RTDETR is adopted.

[0071] Further, see Figure 6 As shown, the RFNet backbone network is an improved version of RFNet. The core architecture is based on ResNet50, replacing all residual blocks with Faster blocks. Faster blocks not only have the residual structure of ResNet, but also incorporate some convolutional PCConv in each Faster block.

[0072] Specifically, this design concept not only significantly reduces computational redundancy but also further optimizes memory access. Specifically, the partial convolution PCConv performs regular convolution operations only on a portion of the input channels to extract spatial features, while keeping the remaining channels unchanged. For continuous or regular memory access, the first or last continuous channel is chosen to be used as a representative of the entire feature map for computation.

[0073] Further, see Figure 8 As shown, the multi-scale convolutional attention mechanism includes the following specific components:

[0074] First, local information is integrated using depthwise convolution techniques;

[0075] Secondly, multi-branch deep strip convolution design is used to obtain multi-scale contextual information;

[0076] Finally, a 1×1 convolution is used to simulate and adjust the correlation between different feature channels. The output is directly used as attention weights to readjust the input features of the multi-scale convolutional attention mechanism. The formula for the multi-scale convolutional attention mechanism is as follows:

[0077]

[0078] Where F represents the input feature, and Att and Out represent the attention matrix and output, respectively. DW-Conv represents element-wise matrix multiplication, DW-Conv represents depthwise convolution, and Scale represents scale. i , i∈{0,1,2,3}, represents the i-th branch.

[0079] Specifically, this invention employs a novel multi-scale convolutional attention mechanism, utilizing simpler and larger kernel convolutions. The multi-scale convolutional attention mechanism is added to the P4 and P5 feature maps in the efficient hybrid encoder, achieving effective aggregation of multi-scale features. These modules utilize kernels of different sizes for convolution operations, thereby capturing feature information at different scales. Through simple element-wise multiplication, the multi-scale convolutional attention mechanism can evoke spatial attention, making the model focus more on important spatial regions. The efficient hybrid encoder with the multi-scale convolutional attention mechanism further improves the performance of object detection. It collects multi-level features from different stages of the network and further extracts global contextual information using a feature pyramid structure. This design enables the model to obtain multi-scale contextual information from local to global, thus achieving better adaptability in spatial and channel dimensions. Furthermore, the multi-scale convolutional attention mechanism also promotes the aggregation of information from low to high levels, enabling the model to more comprehensively understand the content of the input image. See the structural improvements below. Figure 7 As shown.

[0080] Furthermore, in each branch, two depthwise strip convolutions are used.

[0081] Specifically, the kernel sizes used in each branch differ, ranging from 7, 11, to 21. The choice of depthwise striped convolution is primarily based on two important considerations: First, from a computational efficiency perspective, a single pair of 7×1 and 1×7 striped convolutions can approximate a standard 7×7 two-dimensional convolution operation, thus significantly reducing computational complexity. Second, given the prevalence of striped objects like seaweed and sea cucumbers in underwater environments, striped convolution can serve as a powerful auxiliary to mesh convolution, effectively enhancing the extraction capability for these specific striped features.

[0082] In the experimental environment, as shown in Table 2, a Windows 10 operating system was used, equipped with an Intel Core i5-10400F processor and 32GB of RAM. The graphics processing unit (GPU) was an Nvidia GeForce RTX 2080 Ti with 22GB of VRAM. For development, Python 3.9 was chosen, and the PyTorch-GPU 2.0.1 deep learning framework was used for model training. To maximize the GPU's computational performance, CUDA 11.8 and cuDNN v8.9.2.26 were used for acceleration.

[0083] For model selection, models from toolkits such as Ultralytics and mmdetection were primarily used for comparative studies. For the YOLO series models, SGD was selected as the optimizer; for other models, the AdamW optimizer was used, with a learning rate of 0.001 and a weight decay coefficient of 0.05. All models converged, and the training cycle varied depending on the model characteristics. To optimize the training process, an early stopping mechanism was introduced. Training was terminated early if the model performance did not improve within 30 consecutive cycles. Furthermore, each batch contained 8 samples, with 2 data loading worker threads, and mosaic and hybrid training modes were enabled. These optimization measures ensured a smooth and efficient training process.

[0084] Table 2 Experimental Environment Configuration

[0085] software and hardware Version operating system Windows 10 Professional Edition Central Processing Unit (CPU) Intel Core i5-10400F Graphics card (GPU) Nvidia GeForce RTX 2080 Ti Video memory 22GB Memory 32GB programming language Python 3.9 GPU parallel computing platform CUDA 11.8

[0086] Experimental Results and Analysis:

[0087] (1) Comparison experiment between RFNet and mainstream backbone networks

[0088] To explore backbone network model architectures suitable for underwater target detection, a series of mainstream backbone networks were introduced in the experiments, including ResNet50, HGNetv2, SwinTransformer, and MobileViT. Considering the complexity of the underwater environment and the need for model speed, lightweight models such as FasterNet, EMO (EfficientModel), and Mobilenetv3 were also introduced. All models in the experiments used a neck network with an efficient hybrid encoder and a detection head based on RT-DETR.

[0089] As shown in Table 3, the EMO_1M efficient lightweight mobile network models high-resolution feature maps through attention operations, which effectively reduces the number of parameters, but its accuracy is poor and its feature extraction capability is weak. In contrast, Mobilenetv3_small also achieves a lightweight effect and outperforms EMO_1M in accuracy, but it still lags behind the best performance. The introduced SwinTransformer_Tiny uses a variational autoencoder structure and improves the accuracy of object detection by using global and local attention modules, but its performance is mediocre across various dimensions. Similarly, MobileViTv3 combines the features of MobileNet and Vision Transformer, but it did not show any unique advantages in this experiment. Notably, the ResNet50 backbone network performed impressively, with an mAP50 of 84.1% and an mAP50-95 of 48.7%, demonstrating high efficiency and stability, proving its suitability for handling complex underwater tasks. FasterNet_T1 reduces memory access by using partial convolutions, achieving the goal of reducing the number of parameters and computation, but its accuracy remains poor.

[0090] Table 3 shows the backbone network comparison experiments on the URPC2021 dataset.

[0091]

[0092] In contrast, the self-developed RFNet combines the advantages of ResNet50 and FasterNet, achieving the highest accuracy. Compared to the benchmark model RTDETR's HGNetv2, RFNet improves the mAP50-95 metric by 1.7 percentage points, mAP50 by 2.4 percentage points, and recall by 2.1 percentage points.

[0093] (2) Comparison experiment between RFM-DETR and popular detectors

[0094] To verify the performance of the proposed algorithm in underwater target detection, this experiment also compared the performance of single-stage detectors (YOLO series), two-stage detectors (Fster-RCNN, Cascade-RCNN), end-to-end detectors (DETR series), and the underwater target detection method based on spatial feature selection using the FCOS model proposed by Cai Da et al., on the URPC2021 dataset. The superiority of the proposed algorithm RFM-DETR was comprehensively compared using these algorithms. To accelerate the convergence of the RCNN series, improved FCOS, Deformable-DETR, and SSD300 detectors, the weights to be trained were used. The following algorithms converged within their respective training epochs.

[0095] As shown in Table 4, both Cascade-RCNN and Faster-RCNN have relatively low accuracy. Their two-stage processing undoubtedly adds extra computational burden. Similarly, the FCOS underwater target detection method with spatial feature selection is based on pre-training on a large dataset, resulting in slow training convergence and high computational resource consumption. Furthermore, it has poor adaptability to changes in the size and shape of underwater targets, leading to low accuracy. Although SSD300 performs well in terms of computational cost and parameter count, its use of pre-defined fixed-size candidate boxes causes it to miss or falsely detect targets of different scales in complex underwater scenes, resulting in low accuracy.

[0096] Table 4 Comparison of the algorithm of this invention with mainstream detectors on the URPC2021 dataset.

[0097]

[0098]

[0099] In Table 4, Deformable-DETR achieved significant results in accuracy, but still falls short of the best performance. The main reason is its heavy reliance on self-attention mechanisms and deformable sampling, which, while improving accuracy, also increases computational cost, resulting in relatively slow training and inference speeds. This undoubtedly limits its application in real-time underwater target detection. RTDETR performed poorly in accuracy, which is related to the shortcomings of the backbone network HGNetv2 in adapting to underwater environments and the effectiveness of feature extraction (see Experiment Table 3). In contrast, RFM-DETR significantly outperformed RT-DETR and Deformable-DETR. As shown in Table 4, RFM-DETR significantly outperformed the YOLO series models in accuracy, computational cost, and parameter count. For YOLOx, although it uses a deep network structure, it relies too heavily on training data during training, leading to poor performance on new data. For YOLOv6-m and YOLOv5-l, although they also use deep networks, their performance in the feature extraction stage is insufficient, failing to fully extract semantic information, resulting in missed detections of small underwater objects. As for GoldYO11-1, although it introduces an information aggregation-distribution mechanism to enhance information fusion, its excessive number of parameters leads to increased model complexity, lower generalization ability, and suboptimal accuracy. Furthermore, a YOLOv11 detector is introduced. Considering the large x-size is mismatched with the model size of this invention, a smaller YOLOv11-1 size is used for comparison. The algorithm of this invention outperforms YOLOv11-1, with a 1.3 percentage point improvement in mAP50-95 and a 1 percentage point improvement in mAP50. This indicates that the current algorithm's performance is not optimal, which is likely due to the relatively insufficient depth of the YOLOv11-1 model, thus affecting its performance to some extent.

[0100] To more clearly demonstrate the comparison experiment between the algorithm of this invention and mainstream detectors, a scatter plot was drawn with the parameter quantity as the x-axis and mAP50 (%) as the y-axis. See [link to scatter plot]. Figure 9 As shown.

[0101] (3) Ablation test

[0102] To verify the reliability of the algorithm of this invention, ablation experiments are provided. This section will gradually add the designed structure to demonstrate the detailed accuracy changes. The combination of ResNet50 and FasterNet_T1 constitutes the backbone network RFNet proposed in this invention.

[0103] Table 5 Ablation Experiment

[0104]

[0105] As shown in the ablation experiments in Table 5, the algorithm achieved optimal performance after 200 rounds of training using MSCA Attention, ResNet50, and FasterNet_T1 simultaneously, with mAP50-95 reaching 50.5% and mAP50 reaching 85.6%, further demonstrating the effectiveness of the algorithm of this invention in underwater target detection tasks.

[0106] Table 6 Comparison of the algorithm of this invention with other methods on the DUO dataset.

[0107]

[0108] (4) Generalization test

[0109] To verify the model's generalization performance, experiments were also conducted on the DUO underwater dataset. In the experiments, all detector settings used for comparison followed the descriptions in this invention.

[0110] As shown in Table 6, the method of this invention also performs exceptionally well on the DUO underwater dataset, achieving an mAP50-95 of 66.8% and an mAP50 of 86.3%, thus winning with the highest accuracy. Compared to the previous algorithm, there are gains of 0.5% in mAP50 and 0.6% in mAP50-95.

[0111] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0112] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting marine biological targets based on RFM-DETR, characterized in that, Includes the following steps: S1. Obtain a dataset of marine life images and perform enhancement processing on the dataset; S2. Input the seabed organism image dataset into the RFM-DETR target detection algorithm for training to obtain the seabed organism target detection model; S3. Input the image of the seabed organism to be identified into the seabed organism target detection model to obtain the detection result of the seabed organism target to be identified.

2. The method for detecting marine biological targets based on RFM-DETR according to claim 1, characterized in that, The specific steps for enhancing the underwater life image dataset in S1 are as follows: An image enhancement method based on the Transformer diffusion model is used to enhance a dataset of marine life images. The diffusion model includes a forward process and a backward process. The forward process is a Markov process that gradually adds normal Gaussian noise to destroy the image, gradually removing image details until it becomes pure noise. In the backward process, pure noise is input and a neural network is trained to gradually remove the noise introduced by the forward process.

3. The method for detecting marine biological targets based on RFM-DETR according to claim 2, characterized in that, The specific details of the reverse process are as follows: Introducing the conditional diffusion model, in which, given input (x) t (,c,t), where x t The image is a noisy image, c represents the conditional image, t represents the time step, and the denoising network is used to estimate the noise distribution ∈ t During training, the L loss function is used to optimize the denoising network, as shown in the following formula: L s =||∈ t -∈ θ (x t ,c,t)|| Where, ∈ θ The image is a predicted noisy image. The inverse process in the diffusion model is an iterative denoising process or an inverse Markov process, i.e., estimating q(x). t-1 |x t First, use pure Gaussian noise x T The estimation is performed using the following formula: Where, p θ (x 0:T |c) is given the conditional image c, from the noise x T The joint probability distribution of generating a clean image x0 step by step, p(x T ) is the initial Gaussian noise x T The prior distribution of p θ (x t-1 |x t c) is at step t, given the noisy image x t Given the conditional image c, predict the previous image x. t-1 The probability distribution is assumed to be Gaussian, μ θ (x t ,c,t) is the mean, ∑ θ (x t (c,t) represents the variance, and T is the total number of time steps, i.e., the number of time steps required for the diffusion process to go from a pure noise state to a clear image state. It is a Gaussian distribution; When β t When the size is very small, the reverse process of each step is treated as a Gaussian process, and a network is designed to estimate the mean μ. θ (·) and variance ∑ θ (·), the formula for the average value is as follows: By sampling a Gaussian distribution using a reparameterization technique, given a Gaussian distribution... Next, sample x using z = μ + σ⊙∈, where ∈ is a normal Gaussian distribution and ⊙ is the element-wise product; finally, iteratively generate sample images using the following formula until the final image x0 is obtained, as follows:

4. The method for detecting marine biological targets based on RFM-DETR according to claim 1, characterized in that, The RFM-DETR object detection algorithm includes: an RFNet backbone network, an efficient hybrid encoder based on a multi-scale convolutional attention mechanism, and an RTDETR decoder.

5. The method for detecting seabed biological targets based on RFM-DETR according to claim 4, characterized in that, The RFNet backbone network is an improved version of RFNet. Its core architecture is based on ResNet50, replacing all residual blocks with Faster blocks. Faster blocks not only have the residual structure of ResNet, but also incorporate some convolutional PCConv in each Faster block.

6. The method for detecting seabed biological targets based on RFM-DETR according to claim 4, characterized in that, The multi-scale convolutional attention mechanism includes the following specific components: First, local information is integrated using depthwise convolution techniques; Secondly, multi-branch deep strip convolution design is used to obtain multi-scale contextual information; Finally, a 1×1 convolution is used to simulate and adjust the correlation between different feature channels. The output is directly used as attention weights to readjust the input features of the multi-scale convolutional attention mechanism. The formula for the multi-scale convolutional attention mechanism is as follows: Where F represents the input feature, and Att and Out represent the attention matrix and output, respectively. DW-Conv represents element-wise matrix multiplication, DW-Conv represents depthwise convolution, and Scale represents scale. i , i∈{0,1,2,3}, represents the i-th branch.

7. The method for detecting seabed biological targets based on RFM-DETR according to claim 6, characterized in that, In each branch, two depthwise strip convolutions are used.

Images