A jellyfish target detection method based on reinforcement learning

Abstract

The present invention provides a jellyfish target detection method based on reinforcement learning. Compared with the traditional CNN, the Transformer-based feature recognition backbone network TBFRNet of the present invention has better global feature extraction capability; compared with the Transformer, it can improve the regional feature acquisition capability, reduce the computational complexity, and provide module interpretability; based on human perception of the biological characteristics of jellyfish, a jellyfish feature detection module JDM including two submodules FEDM and PDM is designed for the biological characteristics of jellyfish. This new module integrates the human visual system's understanding of the appearance characteristics of jellyfish, and strengthens the network's recognition and extraction capabilities of the jellyfish's umbrella-shaped features and tail features respectively; at the same time, the branch auxiliary detection network BADN can generate gradient information corresponding to the backbone network, thereby playing an auxiliary role in the gradient calculation of the backbone network, helping the deep neural network to retain key features during information transmission, thereby improving the performance and learning efficiency of the model.

Description

Technical Field

[0001] The present invention belongs to the technical fields of image processing, target detection and deep learning, and in particular relates to a jellyfish target detection method based on reinforcement learning. Background Art

[0002] In recent years, driven by factors such as climate change and human activities, jellyfish populations in many oceans around the world have been increasing in number, frequency, and range. Currently, the key environmental factors that trigger jellyfish outbreaks remain unclear, making early warning and intervention difficult to implement. Instead, prevention and control of jellyfish disasters rely primarily on reactive responses after the disaster occurs. This limitation stems largely from a lack of historical data and continuous monitoring methods. Therefore, establishing a real-time jellyfish monitoring system is crucial to improving the effectiveness of jellyfish disaster prevention and control and ensuring the safety of the marine environment. Within this monitoring system, jellyfish detection technology is the most critical component, directly impacting the accuracy and real-time nature of jellyfish monitoring.

[0003] Among existing jellyfish detection technologies, acoustic detection methods can detect the presence of jellyfish, but acoustic image resolution is poor, application costs are high, and jellyfish species identification is difficult. Remote sensing and surface image detection techniques can reveal jellyfish distribution patterns to a certain extent, but they still have limitations in distinguishing jellyfish species and detecting underwater depth distribution. In recent years, detection techniques based on underwater optical images have developed, driven by advances in target detection technology, but generally still represent a transfer of terrestrial detection techniques to underwater environments. Han et al. proposed a jellyfish detection technique that combines image preprocessing and deep learning methods, but this technique has not been validated on real images. Gao et al. proposed a self-supervised jellyfish detection algorithm based on YOLOv5, further expanding the jellyfish dataset and improving jellyfish detection performance. Bi et al. proposed a jellyfish detection technique and jellyfish statistics algorithm based on RCNN and FP16, but only detected moon jellyfish and nebulosa. Zhang et al. proposed the JF-YOLO jellyfish detection algorithm and applied it to real underwater scenes, but the algorithm was unable to accurately report jellyfish species. Based on the YOLOv5 model, Pham et al. achieved high-precision detection of jellyfish by adding a GAM module and replacing the basic convolution. However, the article did not clearly state the detection results and detection types.

[0004] Object detection technology based on underwater optical images has developed rapidly in recent years, particularly with deep learning methods, which have enabled researchers to transfer terrestrial detection techniques to underwater environments. However, current algorithms for jellyfish detection lack targeted optimization and feature extraction strategies for jellyfish's unique biological characteristics. This results in low detection accuracy, difficulty in species identification, and poor identification accuracy. Summary of the Invention

[0005] In order to solve the problems of low detection accuracy, difficult species identification and poor identification accuracy in existing jellyfish detection methods, the present invention provides a jellyfish target detection method based on reinforcement learning, which greatly improves the recognition accuracy and classification accuracy of jellyfish targets in the marine environment.

[0006] A jellyfish target detection method based on reinforcement learning uses a Transformer-based reinforcement learning model to detect jellyfish targets in the image to be tested. The reinforcement learning model includes a feature recognition backbone network, a branch auxiliary detection network, and a detection head.

[0007] The feature recognition backbone network is used to obtain and extract multi-level deep features of the image to be tested, and the last three levels of deep features are recorded as P3 feature map, P4 feature map, and P5 feature map respectively; wherein each level of deep features is extracted based on the previous level of deep features, and the P5 feature map is extracted using a method based on Transformer and independent component analysis;

[0008] The branch-assisted detection network uses dynamic snake convolution to extract the local feature map of the image to be tested, and fuses the local feature map with the P3 feature map, the P4 feature map, and the P5 feature map to obtain the P3' feature map, fuses the P3' feature map with the P4 feature map and the P5 feature map to obtain the P4' feature map, and fuses the P4' feature map with the P5 feature map to obtain the P5' feature map;

[0009] The detection head is used for obtaining the jellyfish target and the category to which the jellyfish target belongs in the image to be detected using the P3 feature map, the P4 feature map, the P5 feature map, the P3' feature map, the P4' feature map and the P5' feature map.

[0010] Furthermore, the feature recognition backbone network includes a STEM unit, a first TABlock unit, a first Downsample unit, a second TABlock unit, a second Downsample unit, a third TABlock unit, a third Downsample unit, and a TABlock+T-ICA fusion unit that are cascaded in sequence;

[0011] The STEM unit is used to perform coarse feature extraction on the image to be tested to obtain a P1 feature map;

[0012] The first TABlock unit is used to extract features from the P1 feature map to obtain the P2 feature map;

[0013] The first Downsample unit is used to downsample the P2 feature map to obtain a downsampled P2 feature map;

[0014] The second TABlock unit is used to extract features from the downsampled P2 feature map to obtain a P3 feature map;

[0015] The second Downsample unit is used to downsample the P3 feature map to obtain a downsampled P3 feature map;

[0016] The third TABlock unit is used to extract features from the downsampled P3 feature map to obtain a P4 feature map;

[0017] The third Downsample unit is used to downsample the P4 feature map to obtain a downsampled P4 feature map;

[0018] The TABlock+T-ICA fusion unit is used to perform deep feature extraction on the downsampled P4 feature map to obtain a P5 feature map.

[0019] Furthermore, any TABlock unit includes a pooling layer, a first convolutional layer, a first BN layer, a normalization layer, a second convolutional layer, and a second BN layer connected in sequence. The method for obtaining any one of the P1 feature map to the P4 feature map is:

[0020] x i+1 =Conv B (Conv B,G (I i ))+I i

[0021] I i =Pool(x i )+x i

[0022] Among them, x i It is the downsampled feature map output by the previous Downsample unit. When obtaining the P1 feature map, x i P1 feature map output by the STEM unit, Conv B,G (·) represents the joint operation of the first convolutional layer, the first BN layer, and the normalization layer. B (·) represents the joint operation of the second convolutional layer and the second BN layer, I i Represents the intermediate feature map of the TABlock unit at this level, x i+1 It represents the feature map output by the TABlock unit at this level, and Pool(·) represents the pooling operation of the pooling layer.

[0023] Furthermore, the T-ICA unit in the TABlock+T-ICA fusion unit includes a first LN layer, a first Linear layer, a WDICA layer, a QKV mapping layer, an MHSA layer, a second Linear layer, a second LN layer, and an MLP perceptron connected in sequence. The method for obtaining the P5 feature map is:

[0024] x i+1 =MLP(LN(I i ))+I i

[0025] I i =Linear(MHSA(WD(Linear(LN(x i ))))+x i )

[0026] Q=WD high (Linear(x i ))

[0027] K=WD low (Linear(x i ))

[0028] V=WD low (Linear(x i ))

[0029]

[0030] Among them, x i+1 is the P5 feature map, x i is the feature map output by the TABlock unit in the TABlock+T-ICA fusion unit, I i is the intermediate feature map in the TABlock+T-ICA fusion unit, LN(·) represents the layer normalization operation provided by any LN layer, MLP(·) represents the multi-layer perception operation provided by the MLP perceptron, Linear(·) represents the linear normalization operation provided by any Linear layer, and WD high (·) represents the wavelet decomposition operation of the high-frequency component, WD low (·) represents the wavelet decomposition operation of the low-frequency component, Q is the component obtained by the wavelet decomposition operation of the high-frequency component, K is the component obtained by the wavelet decomposition operation of the low-frequency component, V is another component obtained by the wavelet decomposition operation of the low-frequency component, MHSA(·) is the multi-head attention operation provided by the MHSA layer, C3 is the number of channels of the third TABlock unit, b is the set bias coefficient, and Softmax(·) is the activation function.

[0031] Furthermore, the feature recognition backbone network also includes a lightweight unit SPPFLSKA, and the P3 feature map, the P4 feature map, and the P5 feature map are respectively fused and lightweight processed by the lightweight unit SPPFLSKA and then output to the subsequent network;

[0032] Among them, the lightweight unit SPPFLSKA includes the first Conv layer, the first Maxpool layer, the second Maxpool layer, the third Maxpool layer, the LSKA layer and the second Conv layer cascaded in sequence; after the features of any input are processed by the first Conv layer, the first Maxpool layer, the second Maxpool layer, the third Maxpool layer and the LSKA layer in sequence, the feature map obtained is superimposed with the feature map output by the first Conv layer, the feature map output by the first Maxpool layer and the feature map output by the second Maxpool layer, and the superimposed feature map is convolved by the second Conv layer to obtain the feature map after the lightweight processing;

[0033] Among them, the LSKA layer is composed of the first small-size convolution layer DWConv, the first small-size convolution layer DWConv, the first large-size convolution layer DWDConv, the second large-size convolution layer DWDConv, and the Conv layer, which are connected in series in sequence. The feature map output by the Conv layer is multiplied by the feature map input to the LSKA layer, and the product is the feature map finally output by the LSKA layer.

[0034] Furthermore, the branch auxiliary detection network includes a feature enhancement detection unit FEDM, a dynamic perception unit DPM, a first CBS unit to a third CBS unit, a first Fuse unit to a third Fuse unit, and a first Conv+DPM fusion unit to a third Conv+DPM fusion unit;

[0035] The feature enhancement detection unit FEDM is used to extract features from the image to be tested by sliding filtering to obtain an intermediate feature map;

[0036] The dynamic perception unit DPM is used to extract local features of the intermediate feature map through dynamic snake convolution to obtain a local feature map;

[0037] The first to third CBS units are used to perform convolution processing on the P3 feature map, the P4 feature map, and the P5 feature map, respectively, to obtain the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map respectively;

[0038] The first Fuse unit is used to fuse the local feature map with the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map to obtain the P3' initial feature map;

[0039] The first Conv+DPM fusion unit is used to extract local features from the P3' initial feature map to obtain the P3' feature map;

[0040] The second Fuse unit is used to fuse the P3' feature map with the P4 convolution feature map and the P5 convolution feature map to obtain the P4' initial feature map;

[0041] The second Conv+DPM fusion unit is used to extract local features from the P4' initial feature map to obtain the P4' feature map;

[0042] The third Fuse unit is used to fuse the P4' feature map and the P5 convolution feature map to obtain the P5' initial feature map;

[0043] The third Conv+DPM fusion unit is used to extract local features from the P5' initial feature map to obtain the P5' feature map.

[0044] Furthermore, the feature enhancement detection unit FEDM includes a basic convolution layer, a feature enhancement convolution layer, and a normalization layer. The method for the feature enhancement detection unit FEDM to obtain the intermediate feature map is as follows:

[0045] The basic convolution layer uses a convolution filter to perform a first convolution operation on the image to be tested to obtain a coarse feature map, wherein the formula of the first convolution operation is as follows:

[0046] I*K1=∑ i ∑ j I(x+i,y+j)·K1(i,j)

[0047] Where I is the image to be tested, K1 is a convolution filter with a kernel size of 3×3, I(x+i,y+j) is the pixel with coordinates (x+i,y+j) on the image to be tested, K1(i,j) is the convolution value of the i-th row and j-th column in the convolution filter K1, and:

[0048]

[0049] The feature enhancement convolution layer uses a sliding filter to perform a second convolution operation on the coarse feature map to obtain an intermediate enhanced feature map Among them, the calculation formula of the second convolution operation is as follows:

[0050]

[0051] in, is a coarse feature map, K2 is a sliding filter with a kernel size of 3×3, is the pixel with coordinates (x+i,y+j) on the coarse feature map, K2(i,j) is the convolution value of the i-th row and j-th column in the sliding filter K2, and:

[0052]

[0053] Among them, each convolution value in the sliding filter K2 satisfies the following constraints:

[0054] K(i,j)≥K(i+m,j+n),n∈N(-1,1),m∈N(-1,1)

[0055] Wherein, N(-1,1) represents a natural number between -1 and 1, m is the horizontal sliding step length, and n is the vertical sliding step length;

[0056] The normalization layer normalizes the intermediate enhanced feature map I to obtain the final intermediate feature map, where the normalization process f(·) is:

[0057]

[0058] Among them, min(·) is the minimum function, max(·) is the maximum function, and γ is the set normalized weight.

[0059] Furthermore, any dynamic perception unit DPM includes CBG unit I, CBG unit II, bottleneck dynamic convolution unit I, bottleneck dynamic convolution unit II, bottleneck dynamic convolution unit III, and Split unit;

[0060] The intermediate feature map output by the feature enhancement detection unit FEDM passes through the CBG unit I, Split unit, and neck dynamic convolution unit I in sequence to obtain the feature Figure I ;

[0061] feature Figure I Enter the bottleneck dynamic convolution unit II and output features Figure II ;

[0062] feature Figure II Enter the bottleneck dynamic convolution unit III and output features Figure III ;

[0063] Combine the feature map output by the Split unit with the feature Figure I ,feature Figure II ,feature Figure III Superposition to obtain a superimposed feature map;

[0064] The superimposed feature map is input into CBG unit II to obtain the final local feature map.

[0065] Furthermore, any bottleneck dynamic convolution unit consists of a dynamic convolution layer and a CBG unit;

[0066] Among them, after any input feature map is processed by the dynamic convolution layer and the CBG unit, the processed feature map is superimposed on the feature map of the original input to obtain the feature map finally output by the current bottleneck dynamic convolution unit.

[0067] Furthermore, any CBG unit is composed of a Conv layer, a BN layer, and a Gelu layer cascaded in sequence.

[0068] Beneficial effects:

[0069] 1. The present invention provides a jellyfish target detection method based on reinforcement learning. Compared with traditional CNN, the Transformer-based feature recognition backbone network TBFRNet of the present invention has better global feature extraction capability. Compared with Transformer, it can improve the regional feature acquisition capability, reduce computational complexity, and provide module interpretability. Based on human perception of the biological characteristics of jellyfish, a jellyfish feature detection module JDM consisting of two submodules FEDM and PDM is designed for the biological characteristics of jellyfish. This new module integrates the human visual system's understanding of the appearance characteristics of jellyfish, and strengthens the network's recognition and extraction capabilities of the jellyfish's umbrella-shaped features and tail features, respectively.

[0070] 2. The present invention provides a jellyfish target detection method based on reinforcement learning. On the basis of the original backbone network, a branch auxiliary detection network BADN is added in parallel. BADN can generate gradient information corresponding to the backbone network, thereby playing an auxiliary role in the gradient calculation of the backbone network, helping the deep neural network to retain key features during information transmission, thereby improving the performance and learning efficiency of the model. Moreover, because it can be removed during the inference stage, it does not bring too much additional computing overhead. This design makes BADN very suitable for real-time detection application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a general principle block diagram of the reinforcement learning model provided by the present invention;

[0072] Figure 2 A diagram of the research ideas provided for the present invention;

[0073] Figure 3 Schematic diagram of the backbone network TBFRNet provided by the present invention;

[0074] Figure 4 A schematic diagram of the lightweight network provided by the present invention;

[0075] Figure 5 Schematic diagram of the TABlock provided by the present invention;

[0076] Figure 6A schematic diagram of the downsampling unit provided by the present invention;

[0077] Figure 7 A schematic diagram of the branch-assisted detection network provided by the present invention;

[0078] Figure 8 Schematic diagram of the dynamic perception unit DPM provided by the present invention;

[0079] Figure 9 This is a schematic diagram of the detection head provided by the present invention. DETAILED DESCRIPTION

[0080] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0081] The key to jellyfish detection is to extract rich, fine-grained information with jellyfish characteristics. However, the types and shapes of jellyfish are more complex than those of other terrestrial and marine organisms, making jellyfish detection more difficult. To address the above problems, this paper proposes a jellyfish target detection algorithm based on reinforcement learning. It uses a Transformer-based reinforcement learning model to detect jellyfish targets in the image to be tested. Figure 1 As shown, the reinforcement learning model includes a feature recognition backbone network, a branch auxiliary detection network and a detection head;

[0082] The feature recognition backbone network is used to obtain and extract multi-level deep features of the image to be tested, and the last three levels of deep features are recorded as P3 feature map, P4 feature map, and P5 feature map respectively; wherein each level of deep features is extracted based on the previous level of deep features, and the P5 feature map is extracted using a method based on Transformer and independent component analysis;

[0083] The branch-assisted detection network uses dynamic snake convolution to extract the local feature map of the image to be tested, and fuses the local feature map with the P3 feature map, the P4 feature map, and the P5 feature map to obtain the P3' feature map, fuses the P3' feature map with the P4 feature map and the P5 feature map to obtain the P4' feature map, and fuses the P4' feature map with the P5 feature map to obtain the P5' feature map;

[0084] The detection head is used for obtaining the jellyfish target and the category to which the jellyfish target belongs in the image to be detected using the P3 feature map, the P4 feature map, the P5 feature map, the P3' feature map, the P4' feature map and the P5' feature map.

[0085] It should be noted that in order to meet the requirements of jellyfish detection and identification, e.g. Figure 2As shown, when training a reinforcement learning model, the present invention constructed a jellyfish image dataset containing over 10,000 jellyfish images, covering a wide range of jellyfish species and in various underwater environments. The dataset was primarily sourced from the following sources: images of jellyfish categories selected from ImageNet, images collected via web crawlers, and images of jellyfish cultured in the laboratory and captured in the field. The present invention divided the jellyfish dataset into a training set and a test set, with a ratio of 80% to 20%. The training set contains 8,356 images, and the test set contains 2,267 images. This division ensures the diversity and balance of the dataset. The present invention collected images of 12 jellyfish categories from different waters, including comb jellies, jellyfish, egg yolk jellyfish, sea thorn jellyfish, spotted jellyfish, lion's mane jellyfish, nitrate jellyfish, golden jellyfish, mosaic jellyfish, lighthouse jellyfish, inverted jellyfish, and moon jellyfish.

[0086] Further, if Figure 3 As shown, the feature recognition backbone network includes a STEM unit, a first TABlock unit, a first Downsample unit, a second TABlock unit, a second Downsample unit, a third TABlock unit, a third Downsample unit, and a TABlock+T-ICA fusion unit that are cascaded in sequence;

[0087] The STEM unit is used to perform coarse feature extraction on the image to be tested to obtain a P1 feature map; the first TABlock unit is used to perform feature extraction on the P1 feature map to obtain a P2 feature map; the first Downsample unit is used to downsample the P2 feature map to obtain a downsampled P2 feature map; the second TABlock unit is used to perform feature extraction on the downsampled P2 feature map to obtain a P3 feature map; the second Downsample unit is used to downsample the P3 feature map to obtain a downsampled P3 feature map; the third TABlock unit is used to perform feature extraction on the downsampled P3 feature map to obtain a P4 feature map; the third Downsample unit is used to downsample the P4 feature map to obtain a downsampled P4 feature map; the TABlock+T-ICA fusion unit is used to perform deep feature extraction on the downsampled P4 feature map to obtain a P5 feature map.

[0088] It should be noted that in order to reduce computational complexity and achieve lightweight operation of the network, the present invention uses a new spatial pyramid pooling module, which combines the characteristics of spatial pyramid pooling (SPP) and large-size separation convolution kernel attention module (LSKA), while achieving the effect of enhancing the receptive field, reducing the computational complexity of the network. Based on this, the feature recognition backbone network of the present invention also includes a lightweight unit SPPFLSKA, and the P3 feature map, P4 feature map, and P5 feature map are respectively fused and lightweight processed by the lightweight unit SPPFLSKA and then output to the subsequent network;

[0089] Among them, such as Figure 4 As shown, the lightweight unit SPPFLSKA includes a first Conv layer, a first Maxpool layer, a second Maxpool layer, a third Maxpool layer, an LSKA layer, and a second Conv layer cascaded in sequence; after any input feature is processed by the first Conv layer, the first Maxpool layer, the second Maxpool layer, the third Maxpool layer, and the LSKA layer in sequence, the obtained feature map is superimposed with the feature map output by the first Conv layer, the feature map output by the first Maxpool layer, and the feature map output by the second Maxpool layer, and the superimposed feature map is then convolved by the second Conv layer to obtain a feature map after the lightweight processing;

[0090] Among them, the LSKA layer is composed of the first small-size convolution layer DWConv, the first small-size convolution layer DWConv, the first large-size convolution layer DWDConv, the second large-size convolution layer DWDConv, and the Conv layer, which are connected in series in sequence. The feature map output by the Conv layer is multiplied by the feature map input to the LSKA layer, and the product is the feature map finally output by the LSKA layer.

[0091] It should be noted that a key advantage of the SPPFLSKA structure is its effectiveness in multi-scale feature aggregation, which not only enables the model to capture richer and more subtle feature information, but also enhances the model's adaptability to changes in image size. In practical applications, fluctuations in image size are common, and the pooling strategy of SPPFLSKA ensures that the output feature vector size is fixed regardless of how the input image size changes. This feature is crucial to improving the stability and generalization ability of the model. In the present invention, SPPFLSKA is placed at the end of the backbone network because it can provide effective processing for input images of different sizes, enhance the diversity and robustness of features, and promote the lightweight model.

[0092] Thus, the present invention proposes a Transformer-based feature recognition backbone network in combination with the Darknet network architecture. The backbone network sets up four stages (four TABlock units) to meet the requirements of feature output.

[0093] First, the input image is processed by a regular convolutional layer, which is two 3×3 convolutions with a stride of 2. Its operation can be regarded as a patch embedding in the Transformer, so it can be expressed as

[0094]

[0095] Among them, x0 is the input, x1 is the output, B is the batch size, H×W represents the image size, C j It is the feature channel width of the jth level, that is, the number of TABlock units. In fact, this can be used as the P1 output layer of Darknet, but generally because this layer cannot provide effective image features, the output of the P1 layer is directly defined as None in the present invention. Thus, the network enters the first stage, where the TABlock module is designed using PoolMixer in combination with the convolutional layer to extract the underlying features. In the use of PoolMixer, a residual module is set to achieve full extraction of image features, and then the convolutional layer is set again for feature extraction, and the convolutional layer is also similarly set with a residual module.

[0096] Further, if Figure 5 As shown, any TABlock unit includes a pooling layer, a first convolutional layer, a first BN layer, a normalization layer, a second convolutional layer, and a second BN layer connected in sequence. The method for obtaining any one of the P1 feature map to the P4 feature map is:

[0097]

[0098] Among them, x i It is the downsampled feature map output by the previous Downsample unit. When obtaining the P1 feature map, x i P1 feature map output by the STEM unit, Conv B,G (·) represents the joint operation of the first convolutional layer, the first BN layer, and the normalization layer GeLU. B (·) represents the joint operation of the second convolutional layer and the second BN layer, I i Represents the intermediate feature map of the TABlock unit at this level, x i+1 It represents the feature map output by the TABlock unit at this level, and Pool(·) represents the pooling operation of the pooling layer.

[0099] After the feature extraction of the TABlock unit is completed, the image is downsampled. Here, the present invention also optimizes the downsampling process. The present invention uses the RepViTBlock module to further deepen the downsampling layer, and uses depthwise separable convolution to increase the speed of the downsampling module. The improved structure is as follows: Figure 6 The present invention obtains the P3 feature map, the P4 feature map, and the P5 feature map through three identical TABlock units, which can reduce the amount of calculation and improve the stability of feature extraction.

[0100] Further, if Figure 3 As shown, the T-ICA unit in the TABlock+T-ICA fusion unit includes the first LN layer, the first Linear layer, the WDICA layer, the QKV mapping layer, the MHSA layer, the second Linear layer, the second LN layer, and the MLP perceptron connected in sequence. The method for obtaining the P5 feature map is:

[0101]

[0102]

[0103] in, is the P5 feature map, which is the deep feature of the image to be tested, x i is the feature map output by the TABlock unit in the TABlock+T-ICA fusion unit, I i is the intermediate feature map in the TABlock+T-ICA fusion unit, LN(·) represents the layer normalization operation provided by any LN layer, MLP(·) represents the multi-layer perception operation provided by the MLP perceptron, Linear(·) represents the linear normalization operation provided by any Linear layer, and WD high (·) represents the wavelet decomposition operation of high-frequency components, high represents high-frequency components, WD low (·) represents the wavelet decomposition operation of the low-frequency component, low represents the low-frequency component, Q is the component obtained by the wavelet decomposition operation of the high-frequency component, K is the component obtained by the wavelet decomposition operation of the low-frequency component, V is another component obtained by the wavelet decomposition operation of the low-frequency component, MHSA(·) is the multi-head attention operation provided by the MHSA layer, C3 is the number of channels of the third TABlock unit, b is the bias coefficient set for the feedback of the interaction between Q and K in learning, and Softmax(·) is the activation function.

[0104] The Linear layer is the basic layer of the neural network, and its operation can be expressed as

[0105] y=Linear(x)=Wx+b

[0106] Among them, x is the input vector, y is the output vector, and W is the weight matrix, which can be changed through learning.

[0107] It should be noted that in the traditional Transformer, the image needs to be divided into tiles of 16x16 pixels in size, and then each tile is flattened into a one-dimensional vector. Next, linear projection is performed to map the features of the image at different stages to Q (Query), K (Key) and V (Value) channels, and these vectors are used to calculate the correlation between different patches in the image, thereby capturing the high-level semantic information of the image. This process is one of the key steps for Transformer to process images. However, this way of mapping features to channels also brings about problems such as poor regional feature detection effect and poor interpretability. Therefore, the present invention proposes a deep feature extraction model based on Transformer and independent component analysis, named T-ICA.

[0108] An analysis of the Transformer's attention mechanism reveals that the Q channel represents the information each tile wishes to focus on. The K channel primarily extracts similarity information with other tiles Q. The V channel stores the original tile information. Once Q and K determine the association between tiles, the V parameter participates in the weighted summation to form a new feature representation. A deeper analysis of the information represented by these three parameters reveals that the Q channel represents the image's features, K represents the coefficient vector for region-related correlations, and V represents the image weight. Therefore, the present invention innovatively utilizes independent component analysis to calculate the Q, K, and V channels. Based on the principles of digital image processing, it is known that the high-frequency components of an image reflect its texture features, while the low-frequency components reflect its global features. Therefore, the present invention utilizes a frequency-domain-based independent component separation method, setting the Q component to a high-frequency component to reflect the image's features, while K and V are both low-frequency components. V barely participates in the learning cycle, effectively preserving the original image's features as fusion weights. The low-frequency components of K are used as global features to calculate correlations with the high-frequency components of other tiles. Through this frequency domain independent component analysis method, we can cleverly solve the scientifically unexplainable problems in the original Transformer and improve the regional feature learning effect.

[0109] It can be seen that the output of the P5 layer in the Darknet architecture is completed through the T-ICA model proposed in the present invention. The Transformer-based feature recognition backbone network TBFRNet proposed in the present invention has better global feature extraction capabilities than traditional CNN; compared with Transformer, it can improve the regional feature acquisition capability, reduce computational complexity, and provide module interpretability. At the same time, the present invention names the combination of independent component analysis and multi-head attention mechanism in T-ICA as the FastICA attention mechanism, which can be applied to other networks as an attention mechanism.

[0110] It should be noted that since the present invention is aimed at detecting jellyfish in the ocean and has a clear detection target, in order to improve the detection efficiency and accuracy of jellyfish, the present invention designs a jellyfish feature detection module JDM composed of FEDM and DPM based on the biological characteristics of jellyfish. This new module integrates the human visual system's understanding of the appearance characteristics of jellyfish, aiming to enhance the feature recognition ability of the jellyfish cap and tail.

[0111] Specifically, such as Figure 7 As shown, the branch auxiliary detection network includes a feature enhancement detection unit FEDM, a dynamic perception unit DPM, a first CBS unit to a third CBS unit, a first Fuse unit to a third Fuse unit, and a first Conv+DPM fusion unit to a third Conv+DPM fusion unit;

[0112] The feature enhancement detection unit FEDM is used to extract features of the image to be tested by sliding filtering to obtain an intermediate feature map; the dynamic perception unit DPM is used to extract local features of the intermediate feature map by dynamic snake convolution to obtain a local feature map; the first CBS unit to the third CBS unit are used to perform convolution processing on the P3 feature map, the P4 feature map, and the P5 feature map, respectively, to obtain the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map; the first Fuse unit is used to fuse the local feature map with the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map to obtain the P3' initial feature map; the first The Conv+DPM fusion unit is used to perform local feature extraction on the P3' initial feature map to obtain the P3' feature map; the second Fuse unit is used to fuse the P3' feature map with the P4 convolution feature map and the P5 convolution feature map to obtain the P4' initial feature map; the second Conv+DPM fusion unit is used to perform local feature extraction on the P4' initial feature map to obtain the P4' feature map; the third Fuse unit is used to fuse the P4' feature map with the P5 convolution feature map to obtain the P5' initial feature map; the third Conv+DPM fusion unit is used to perform local feature extraction on the P5' initial feature map to obtain the P5' feature map.

[0113] First, we designed a Feature Enhancement Detection Module (FEDM). In the human visual perception system, jellyfish are primarily distinguished from other organisms by their umbrella-shaped features. Therefore, we designed the FEDM to enhance learning of the circular umbrella-shaped features of jellyfish. The FEDM consists of two convolution layers: a base convolution and a feature enhancement convolution. The convolution operation applies a sliding filter to the input image and calculates the dot product between the filter and the input image patch at each position.

[0114] Based on this, the feature enhancement detection unit FEDM includes a basic convolution layer, a feature enhancement convolution layer and a normalization layer. The method for the feature enhancement detection unit FEDM to obtain the intermediate feature map is as follows:

[0115] The basic convolution layer uses a convolution filter to perform a first convolution operation on the image to be tested to obtain a coarse feature map, wherein the formula of the first convolution operation is as follows:

[0116] I*K1=∑ i ∑ j I(x+i,y+j)·K1(i,j)

[0117] Where I is the image to be tested, K1 is a convolution filter with a kernel size of 3×3, I(x+i,y+j) is the pixel with coordinates (x+i,y+j) on the image to be tested, K1(i,j) is the convolution value of the i-th row and j-th column in the convolution filter K1, and:

[0118]

[0119] The feature enhancement convolution layer uses a sliding filter to perform a second convolution operation on the coarse feature map to obtain an intermediate enhanced feature map Among them, the calculation formula of the second convolution operation is as follows:

[0120]

[0121] in, is a coarse feature map, K2 is a sliding filter with a kernel size of 3×3, is the pixel with coordinates (x+i,y+j) on the coarse feature map, K2(i,j) is the convolution value of the i-th row and j-th column in the sliding filter K2, and:

[0122]

[0123] Among them, each convolution value in the sliding filter K2 satisfies the following constraints:

[0124] K(i,j)≥K(i+m,j+n),n∈N(-1,1),m∈N(-1,1)

[0125] Wherein, N(-1,1) represents a natural number between -1 and 1, m is the horizontal sliding step length, and n is the vertical sliding step length;

[0126] Specifically, the first layer of the FEDM architecture employs a set of 3×3 convolution kernels, which extract basic features from the input image. To enhance the convolution layer's sensitivity to circular features, a second feature-enhancing convolution layer is added. This layer also employs a set of 3×3 convolution kernels, acting on the output of the first convolution layer. In this layer, convolution weights are constrained, increasing the weight of the center.

[0127] The purpose of this design is to enable the convolution kernel to achieve better results in detecting edge features, which is especially important for the recognition of circular features. Through the collaboration of these two layers of convolution, the circular features of the jellyfish and the species characteristics of the circular umbrella inside can be better extracted, and the output feature image is normalized and enhanced.

[0128] Furthermore, the normalization layer enhances the intermediate feature map Perform normalization to obtain the final intermediate feature map, where the normalization process f(·) is:

[0129]

[0130] Among them, min(·) is the minimum function, max(·) is the maximum function, γ is the set normalized weight, is the final intermediate feature map.

[0131] Secondly, the present invention designs a dynamic perception module (DPM). Because jellyfish have a high degree of freedom in water, their morphology in images is very complex. Furthermore, the tail features of some jellyfish species are more prominent than their umbrella-shaped structures. While this tail feature can serve as an identification feature for jellyfish, it can also cause overdetection and identification errors during actual detection. This complexity and error must be considered and addressed in the design of jellyfish target detection algorithms to ensure accuracy and robustness. To this end, the present invention designs a dynamic perception module (DPM) that incorporates dynamic serpentine convolution. Dynamic serpentine convolution can adaptively focus on slender and winding local structures to accurately capture the characteristics of tubular structures. The core concept of this convolution method is to enhance perception capabilities through dynamically shaped convolution kernels, optimizing feature extraction for tubular structures. This feature extraction concept aligns with the characteristics of jellyfish tails encountered in jellyfish detection.

[0132] Specifically, such as Figure 8As shown, any dynamic perception unit DPM includes CBG unit I, CBG unit II, bottleneck dynamic convolution unit I, bottleneck dynamic convolution unit II, bottleneck dynamic convolution unit III, and Split unit;

[0133] The intermediate feature map output by the feature enhancement detection unit FEDM passes through the CBG unit I, Split unit, and neck dynamic convolution unit I in sequence to obtain the feature Figure I ;feature Figure I Enter the bottleneck dynamic convolution unit II and output features Figure II ;feature Figure II Enter the bottleneck dynamic convolution unit III and output features Figure III ; Combine the feature map output by the Split unit with the feature Figure I ,feature Figure II ,feature Figure III The superimposed feature map is input into the CBG unit II to obtain the final local feature map.

[0134] Specifically, any bottleneck dynamic convolution unit is composed of a dynamic convolution layer and a CBG unit; wherein, after any input feature map is processed by the dynamic convolution layer and the CBG unit, the processed feature map is superimposed on the original input feature map to obtain the final output feature map of the current bottleneck dynamic convolution unit.

[0135] Furthermore, any CBG unit is composed of a Conv layer, a BN layer, and a Gelu layer cascaded in sequence.

[0136] As can be seen, in the DPM, the present invention first employs a convolutional layer consisting of 1×1 convolutional kernels. This layer is functionally equivalent to a fully connected layer, introducing more nonlinear features while maintaining the input size, thereby enhancing the neural network's expressive power. Next, the module transitions to a three-layer dynamic feature extraction layer, led by DysnakeConv. Each of these layers consists of a 3×3 dynamic snake convolution followed by a traditional 3×3 convolutional layer. This design aims to capture and extract the dynamic characteristics of jellyfish in underwater images through DysnakeConv. Following these three dynamic feature extraction layers, a cross-layer linking mechanism further enhances the ability to extract dynamic jellyfish features. Finally, a 1×1 convolutional layer is used again to achieve full connectivity, transforming the extracted feature map into the final class prediction. This module design combines traditional convolution operations with a dynamic convolutional architecture. Through multi-scale feature extraction and effective information fusion, it can effectively identify and detect jellyfish with diverse body shapes and features in underwater environments, improving the accuracy and robustness of jellyfish target detection.

[0137] It should be noted that in the traditional Darknet network framework, image information is transmitted downward layer by layer. Information may be lost during this process, resulting in gradient deviation and affecting the update of model parameters. Therefore, this paper proposes a branch-assisted detection network (BADN) to help the network retain the key features required for performing object detection tasks.

[0138] BADN is a network structure that is added in parallel to the original backbone network. BADN also extracts features from the original image through convolutional layers. It can generate gradient information corresponding to the main branch, thereby playing an auxiliary role in the gradient calculation of the main branch. It has two advantages: (1) BADN helps deep neural networks retain key features during information transfer, thereby improving model performance. (2) The reliable gradients that BADN can also generate help optimize model parameters and improve model learning efficiency.

[0139] However, in order for BADN to provide a positive effect, two requirements must be met: (1) its computational complexity should be as low as possible to avoid a large increase in the number of network parameters and computational complexity. (2) it must be able to extract effective positive features of the image. These features must be similar to those of the backbone network to a certain extent. Otherwise, conflicts with the main network may occur at the output, reducing the detection effect of the target. Therefore, in order to avoid excessive computational complexity and improve the effectiveness of jellyfish detection, BADN uses the jellyfish feature detection module (JDM) as the main feature extraction method. In order to extract effective positive gain feature information, the present invention redesigns the network architecture of BADN and introduces the feature output of the backbone network into the branch auxiliary detection network.

[0140] First, the input of the BADN is consistent with the backbone network. Subsequently, the original image is subjected to feature extraction using FEDM and DPM. The present invention defines the extracted features as local feature maps. To efficiently extract features and avoid conflicts with the backbone network, the present invention uses the P3, P4, and P5 feature maps calculated in TBFRNet as feature layers and fuses them into the BADN. This design ensures that the BADN maintains a certain degree of similarity with the features extracted by the backbone network during feature extraction. Within the network, the present invention utilizes a CBS (Convolutional Layer, Batch Normalization Layer, and SiLU Activation Function) module. This module enhances information flow through reversible connections when constructing the network's basic units. The network fuses the local feature maps relearned from the original image using the FEDM and DPM modules with the CBS-processed P3, P4, and P5 feature maps. These fused features are then processed by the DPM module, outputting the BADN's P3' feature layer. Next, P3' is fused with the CBS-processed P4 and P5 feature maps, and further processed by the DPM module, outputting the BADN's P4' feature layer. Finally, the P4' feature map is fused with the P5 feature map processed by CBS and processed by the DPM module to output the P5' feature layer of the BADN. After the above processing, the BADN can generate the P3', P4', and P5' feature layers corresponding to the backbone network. This design enables the auxiliary network to assist the backbone network in better object detection, especially when dealing with objects of different scales, providing more feature information, thereby improving detection accuracy and performance.

[0141] At the same time, in order to achieve effective utilization of the auxiliary detection features P3', P4' and P5', the present invention uses an auxiliary detection head applied to the output end of BADN to assist the main detection head in detection. The auxiliary detection head can capture areas that the main detection head cannot cover, reduce the missed detection rate and thus improve the detection accuracy. The auxiliary detection head works together with the main detection head to more comprehensively identify and locate the target. The present invention sets P3', P4' and P5' obtained by BADN processing to enter the auxiliary detection head, and P3, P4 and P5 obtained in the main network TBFRNet are input into the main detection head. The network framework of BADN working in cooperation with the main network is as follows: Figure 9 shown. Figure 9 In the example, the Conv_cls convolutional layer generates classification predictions, and the Conv_reg convolutional layer generates bounding boxes for regression predictions. Through this designed architecture and modular application, BADN not only improves jellyfish detection performance, but also, because it can be removed during the inference phase, it does not incur excessive additional computational overhead. This design makes BADN well-suited for real-time detection applications.

[0142] The objective evaluation results of the present invention, the existing target detection method and the jellyfish detection algorithm in the jellyfish detection task are shown in Table 1.

[0143] Table 1

[0144]

[0145]

[0146] In summary, existing jellyfish detection methods suffer from low detection accuracy, difficulty in species identification, and poor identification accuracy. This paper proposes a jellyfish target detection method based on reinforcement learning. First, a jellyfish dataset is constructed to provide a foundation for network training. Second, a Transformer-based backbone network, TBFRNet, is constructed to improve jellyfish detection accuracy at a macro level. Third, based on the biological characteristics of jellyfish, a jellyfish feature detection module, JDM, is designed to match human perception and enhance jellyfish feature extraction. Finally, a branch-assisted detection network, BADN, is designed and applied in conjunction with JDM to further improve jellyfish detection accuracy. Finally, the network is lightweighted, and a lightweight module, SPPFLSKA, is constructed to increase the processing speed of the jellyfish target detection algorithm. That is to say, the present invention proposes a T-ICA feature extraction module, and based on this module proposes a feature recognition backbone network TBFRNet; then, based on human perception of the biological characteristics of jellyfish, a jellyfish feature detection module JDM is designed, which includes two sub-modules FEDM and PDM, which respectively improve the extraction capabilities of jellyfish umbrella features and tail features; finally, a branch auxiliary network BADN is proposed and applied in combination with JDM to assist the backbone network in jellyfish target detection, effectively improving the algorithm's jellyfish detection effect.

[0147] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may of course make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A jellyfish target detection method based on reinforcement learning, characterized in that: A Transformer-based reinforcement learning model is used to detect jellyfish targets in the image to be tested, wherein the reinforcement learning model includes a feature recognition backbone network, a branch auxiliary detection network, and a detection head; The feature recognition backbone network is used to obtain and extract multi-level deep features of the image to be tested, and the last three levels of deep features are recorded as P3 feature map, P4 feature map, and P5 feature map respectively; wherein each level of deep features is extracted based on the previous level of deep features, and the P5 feature map is extracted using a method based on Transformer and independent component analysis; The branch-assisted detection network uses dynamic snake convolution to extract the local feature map of the image to be tested, and fuses the local feature map with the P3 feature map, the P4 feature map, and the P5 feature map to obtain the P3' feature map, fuses the P3' feature map with the P4 feature map and the P5 feature map to obtain the P4' feature map, and fuses the P4' feature map with the P5 feature map to obtain the P5' feature map; The detection head is used to obtain the jellyfish target and the category of the jellyfish target in the image to be tested using the P3 feature map, the P4 feature map, the P5 feature map, the P3' feature map, the P4' feature map and the P5' feature map; The feature recognition backbone network includes a STEM unit, a first TABlock unit, a first Downsample unit, a second TABlock unit, a second Downsample unit, a third TABlock unit, a third Downsample unit and a TABlock+T-ICA fusion unit which are cascaded in sequence; The STEM unit is used to perform coarse feature extraction on the image to be tested to obtain a P1 feature map; The first TABlock unit is used to extract features from the P1 feature map to obtain the P2 feature map; The first Downsample unit is used to downsample the P2 feature map to obtain a downsampled P2 feature map; The second TABlock unit is used to extract features from the downsampled P2 feature map to obtain a P3 feature map; The second Downsample unit is used to downsample the P3 feature map to obtain a downsampled P3 feature map; The third TABlock unit is used to extract features from the downsampled P3 feature map to obtain a P4 feature map; The third Downsample unit is used to downsample the P4 feature map to obtain a downsampled P4 feature map; The TABlock+T-ICA fusion unit is used to perform deep feature extraction on the downsampled P4 feature map to obtain a P5 feature map; The T-ICA unit in the TABlock+T-ICA fusion unit includes the first LN layer, the first Linear layer, the WDICA layer, the QKV mapping layer, the MHSA layer, the second Linear layer, the second LN layer, and the MLP perceptron connected in sequence. The method for obtaining the P5 feature map is: in, is the P5 feature map, It is the feature map output by the TABlock unit in the TABlock+T-ICA fusion unit. is the intermediate feature map in the TABlock+T-ICA fusion unit, represents the layer normalization operation provided by any LN layer, Represents the multi-layer perception operation provided by the MLP perceptron, Represents the linear normalization operation provided by any Linear layer, represents the wavelet decomposition operation of the high-frequency component, represents the wavelet decomposition operation of the low-frequency component, is the component obtained by wavelet decomposition of high-frequency components, is the component obtained by wavelet decomposition of the low-frequency component, is another component obtained by wavelet decomposition of the low-frequency component. The multi-head attention operation provided by the MHSA layer, is the number of channels of the third TABlock unit, is the set bias coefficient, is the activation function.

2. The jellyfish target detection method based on reinforcement learning according to claim 1, characterized in that: Any TABlock unit includes the pooling layer, the first convolution layer, the first BN layer, the normalization layer, the second convolution layer, and the second BN layer connected in sequence. The method for obtaining any one of the P1 feature map to the P4 feature map is: in, It is the downsampled feature map output by the previous Downsample unit. When obtaining the P1 feature map, is the P1 feature map output by the STEM unit, Represents the joint operation of the first convolutional layer, the first BN layer, and the normalization layer. Represents the joint operation of the second convolutional layer and the second BN layer, Represents the intermediate feature map of the TABlock unit at this level, Represents the feature map output by the TABlock unit at this level, Represents the pooling operation of the pooling layer.

3. The jellyfish target detection method based on reinforcement learning according to claim 1, characterized in that: The feature recognition backbone network also includes a lightweight unit SPPFLSKA, and the P3 feature map, the P4 feature map, and the P5 feature map are respectively fused and lightweight processed by the lightweight unit SPPFLSKA and then output to the subsequent network; Among them, the lightweight unit SPPFLSKA includes the first Conv layer, the first Maxpool layer, the second Maxpool layer, the third Maxpool layer, the LSKA layer and the second Conv layer cascaded in sequence; after the features of any input are processed by the first Conv layer, the first Maxpool layer, the second Maxpool layer, the third Maxpool layer and the LSKA layer in sequence, the feature map obtained is superimposed with the feature map output by the first Conv layer, the feature map output by the first Maxpool layer and the feature map output by the second Maxpool layer, and the superimposed feature map is then convolved by the second Conv layer to obtain the feature map after the lightweight processing; Among them, the LSKA layer is composed of the first small-size convolution layer DWConv, the first small-size convolution layer DWConv, the first large-size convolution layer DWDConv, the second large-size convolution layer DWDConv, and the Conv layer, which are connected in series in sequence. The feature map output by the Conv layer is multiplied by the feature map input to the LSKA layer, and the product is the feature map finally output by the LSKA layer.

4. The jellyfish target detection method based on reinforcement learning according to claim 1, characterized in that: The branch auxiliary detection network includes a feature enhancement detection unit FEDM, a dynamic perception unit DPM, a first CBS unit to a third CBS unit, a first Fuse unit to a third Fuse unit, and a first Conv+DPM fusion unit to a third Conv+DPM fusion unit; The feature enhancement detection unit FEDM is used to extract features from the image to be tested by sliding filtering to obtain an intermediate feature map; The dynamic perception unit DPM is used to extract local features of the intermediate feature map through dynamic snake convolution to obtain a local feature map; The first to third CBS units are used to perform convolution processing on the P3 feature map, the P4 feature map, and the P5 feature map, respectively, to obtain the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map; The first Fuse unit is used to fuse the local feature map with the P3 convolution feature map, the P4 convolution feature map, and the P5 convolution feature map to obtain the P3' initial feature map; The first Conv+DPM fusion unit is used to extract local features from the P3' initial feature map to obtain the P3' feature map; The second Fuse unit is used to fuse the P3' feature map with the P4 convolution feature map and the P5 convolution feature map to obtain the P4' initial feature map; The second Conv+DPM fusion unit is used to extract local features from the P4' initial feature map to obtain the P4' feature map; The third Fuse unit is used to fuse the P4' feature map and the P5 convolution feature map to obtain the P5' initial feature map; The third Conv+DPM fusion unit is used to extract local features from the P5' initial feature map to obtain the P5' feature map.

5. The jellyfish target detection method based on reinforcement learning according to claim 4, characterized in that: The feature enhancement detection unit FEDM includes a basic convolution layer, a feature enhancement convolution layer, and a normalization layer. The method for the feature enhancement detection unit FEDM to obtain the intermediate feature map is as follows: The basic convolution layer uses a convolution filter to perform a first convolution operation on the image to be tested to obtain a coarse feature map, wherein the formula of the first convolution operation is as follows: in, I is the image to be tested, is a convolution filter with a kernel size of 3×3, The coordinates on the image to be measured are Pixels, is the convolution filter Middle i Rank j The convolution value of the column, and there are: The feature enhancement convolution layer uses a sliding filter to perform a second convolution operation on the coarse feature map to obtain an intermediate enhanced feature map , where the calculation formula for the second convolution operation is as follows: in, is the coarse feature map, is a sliding filter with a kernel size of 3×3, The coordinates on the coarse feature map are Pixels, Sliding filter Middle i Rank j The convolution value of the column, and there are: Among them, the sliding filter Each convolution value in satisfies the following constraints: in, express A natural number between 1 and 2, is the horizontal sliding step length, is the longitudinal sliding step length; The normalization layer enhances the intermediate feature map Perform normalization to obtain the final intermediate feature map, where normalization for: in, is the minimum function, is the maximum value function, is the normalized weight set.

6. The jellyfish target detection method based on reinforcement learning according to claim 4, characterized in that: Any dynamic perception unit DPM includes CBG unit I, CBG unit II, bottleneck dynamic convolution unit I, bottleneck dynamic convolution unit II, bottleneck dynamic convolution unit III, and Split unit; The intermediate feature map output by the feature enhancement detection unit FEDM passes through the CBG unit I, the Split unit, and the neck dynamic convolution unit I in sequence to obtain the feature map I; Feature map I enters the bottleneck dynamic convolution unit II and outputs feature map II; Feature map II enters the bottleneck dynamic convolution unit III and outputs feature map III; The feature map output by the Split unit is superimposed with feature map I, feature map II, and feature map III to obtain a superimposed feature map; The superimposed feature map is input into CBG unit II to obtain the final local feature map.

7. The jellyfish target detection method based on reinforcement learning according to claim 6, characterized in that: Any bottleneck dynamic convolution unit consists of a dynamic convolution layer and a CBG unit; Among them, after any input feature map is processed by the dynamic convolution layer and the CBG unit, the processed feature map is superimposed on the feature map of the original input to obtain the feature map finally output by the current bottleneck dynamic convolution unit.

8. A jellyfish target detection method based on reinforcement learning according to claim 6 or 7, characterized in that: Any CBG unit is composed of a sequentially cascaded Conv layer, BN layer, and Gelu layer.

Images