Underwater sound image transmission method combining feature extraction and semantic compression

Abstract

The invention discloses an underwater acoustic image transmission method combining feature extraction and semantic compression, which relates to the technical field of data transmission, and comprises the following steps: a sending end performs semantic feature extraction on an original underwater acoustic image to obtain a semantic feature set; performing compressed encoding on the semantic feature set to generate compressed semantic data; transmitting the compressed semantic data to a receiving end through an underwater acoustic channel; and the receiving end decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task identification based on the semantic feature set. The method comprises the following steps: performing semantic feature extraction on an underwater acoustic image through a sending end, and converting pixel data into a semantic feature set; the data volume is significantly reduced through compression coding, so that the method adapts to the limited bandwidth of an underwater acoustic channel; and the receiving end performs image reconstruction or direct task identification based on the restored semantic features. According to the method, efficient and reliable transmission of image data in a low-bandwidth and high-packet-loss underwater acoustic environment is realized, and effective maintenance and task availability of semantic information are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data transmission, in particular to a method for underwater acoustic image transmission combining feature extraction and semantic compression. BACKGROUND

[0002] With the deepening of ocean science research, deep-sea resource exploration and underwater monitoring applications, higher requirements are put forward for real-time acquisition and reliable transmission of underwater image data. At present, with the help of sonar imaging and underwater optical imaging technologies, high-value image information containing target objects, topography and environmental features can be collected in complex marine environments. However, as the main medium for underwater information transmission, the underwater acoustic channel has very special physical characteristics, which are extremely limited in available bandwidth, significant signal propagation delay, and high error rate and packet loss rate.

[0003] The current common underwater image transmission methods can be mainly divided into two categories. One is based on traditional image compression coding, such as JPEG, JPEG2000 and other static image compression standards, or H.264 and other video encoding standards. The core idea of this method is to compress the image pixels or bit stream to reduce the data volume. However, this method is not fully optimized for the communication conditions of extremely limited bandwidth, frequent error and packet loss of underwater acoustic channel, resulting in low transmission efficiency in actual underwater environment. More importantly, this method is extremely sensitive to data loss. Once data packet loss occurs during transmission, the quality of the reconstructed image will be severely damaged, and even lose its usability.

[0004] Another method attempts to transmit image information in layers or batches through downsampling or progressive encoding. Although this method reduces the data volume of a single transmission to some extent, it still follows the pixel-centered transmission idea and fails to eliminate redundant information in the image fundamentally, so it is still difficult to achieve efficient and robust image transmission in low-bandwidth underwater acoustic environment. SUMMARY

[0005] The embodiment of the present application provides a method for underwater acoustic image transmission combining feature extraction and semantic compression, which aims to solve the technical problem of how to provide a transmission method that can effectively adapt to the characteristics of underwater acoustic channel, still achieve efficient and reliable image transmission in a limited bandwidth and high packet loss environment, and maintain the usability of its semantic content.

[0006] In a first aspect, the embodiment of the present application provides a method for underwater acoustic image transmission combining feature extraction and semantic compression, which comprises:

[0007] The sending end extracts semantic features from the original underwater acoustic image, and obtains a semantic feature set containing target category information, target position information and scene semantics;

[0008] compressing and encoding the semantic feature set to generate compressed semantic data;

[0009] transmitting the compressed semantic data to a receiving end through an underwater acoustic channel;

[0010] decoding the received compressed semantic data, restoring the semantic feature set, and performing image reconstruction and / or task recognition based on the semantic feature set.

[0011] Further technical solutions are that the semantic feature extraction from the original underwater acoustic image to obtain the semantic feature set containing target category information, target position information, and scene semantics includes:

[0012] preprocessing the original underwater acoustic image to obtain a preprocessed image, the preprocessing including filtering and noise reduction and image enhancement;

[0013] performing multi-level feature analysis on the preprocessed image by using a preset deep learning network to extract multi-level features including global feature vectors, local feature vectors, target category information, and target position information;

[0014] fusing the extracted multi-level features to form the semantic feature set.

[0015] Further technical solutions are that the multi-level feature analysis includes:

[0016] extracting global feature vectors representing scene-level background semantics and overall layout in the preprocessed image;

[0017] extracting target category information and target position information of key targets in the preprocessed image, the target position information including a bounding box and / or a segmentation mask;

[0018] extracting local feature vectors representing texture, edge, and shape inside a target region in the preprocessed image.

[0019] Further technical solutions are that before the compression and encoding of the semantic feature set to generate compressed semantic data, the method further includes:

[0020] performing importance evaluation on features in the semantic feature set;

[0021] based on the importance evaluation result, filtering and retaining task-critical semantic features from the semantic feature set and discarding redundant features.

[0022] Further technical solutions are that the compression and encoding of the semantic feature set to generate compressed semantic data include:

[0023] The target category information is encoded by using a hierarchical dictionary;

[0024] The bounding box in the target position information is encoded by using a difference encoding, and the segmentation mask in the target position information is compressed by using a combination of binarization and run-length encoding and block-based hash encoding;

[0025] The global feature vector is encoded by using vector quantization and entropy encoding;

[0026] The local feature vector is compressed by using low-rank matrix approximation or clustering.

[0027] A further technical solution is that the semantic feature set is compressed and encoded to generate compressed semantic data, and the method further comprises:

[0028] The semantic feature set is divided into a core layer and an enhancement layer, wherein the core layer at least includes target category information, target position information and a global feature vector, and the enhancement layer includes a local feature vector;

[0029] The core layer and the enhancement layer are respectively encoded and packaged;

[0030] The ratio of the compression encoding is dynamically adjusted according to the bandwidth condition of the current underwater acoustic channel, wherein when the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold, only the compressed semantic data of the core layer is transmitted.

[0031] A further technical solution is that the compressed semantic data is transmitted to a receiving end through an underwater acoustic channel, and the method comprises:

[0032] The compressed semantic data is channel-encoded by using an unequal error protection mechanism, wherein the core layer data is allocated a higher redundancy than the enhancement layer data;

[0033] The channel encoding of the compressed semantic data by using the unequal error protection mechanism comprises:

[0034] An extended window fountain code is used to allocate a higher extraction probability to the data of the core layer when the symbol is generated, and a progressive window extension strategy is used, and the initial window preferentially covers the data of the core layer.

[0035] A further technical solution is that the channel encoding adopts a joint source-channel encoding strategy, wherein:

[0036] The core layer data is encoded by using a digital fountain code;

[0037] The enhancement layer data is encoded by using a joint source-channel encoding similar to simulation, and continuous features are embedded into a modulation symbol space.

[0038] Further, the technical scheme is that the decoding of the received compressed semantic data comprises:

[0039] The data of the core layer is preferentially decoded to restore target category information, target position information and a global feature vector.

[0040] On the basis of successful decoding of the data of the core layer, the data of the enhancement layer is continuously decoded in a progressive manner to restore a local feature vector.

[0041] Further, the technical scheme is that the image reconstruction and / or task recognition based on the semantic feature set comprises:

[0042] When only the data of the core layer is successfully decoded, the restored target category information, target position information and global feature vector are directly used to output a task recognition result.

[0043] When the data of the core layer and the enhancement layer is successfully decoded, the original underwater image is reconstructed based on the complete semantic feature set by using a generative model, and a task recognition result is output.

[0044] In a second aspect, an embodiment of the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above method when executing the computer program.

[0045] In a third aspect, an embodiment of the present application further provides a computer readable storage medium, wherein the storage medium stores a computer program, and the computer program can implement the above method when executed by a processor.

[0046] An embodiment of the present application provides an underwater image transmission method combining feature extraction and semantic compression. The method comprises: a sending end performing semantic feature extraction on an original underwater image to obtain a semantic feature set comprising target category information, target position information and scene semantics; performing compression encoding on the semantic feature set to generate compressed semantic data; transmitting the compressed semantic data to a receiving end through an underwater acoustic channel; and the receiving end decoding the received compressed semantic data, restoring the semantic feature set, and performing image reconstruction and / or task recognition based on the semantic feature set. The present application converts pixel data into a semantic feature set comprising target category, position and scene semantics by performing semantic feature extraction on an underwater image at the sending end; then significantly reduces the data volume through compression encoding to adapt to the limited bandwidth of the underwater acoustic channel; and performs image reconstruction or direct task recognition based on the restored semantic features at the receiving end. The method realizes efficient and reliable transmission of image data in a low-bandwidth and high-packet-loss underwater acoustic environment, and ensures effective maintenance of semantic information and task usability. BRIEF DESCRIPTION OF DRAWINGS

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating an underwater acoustic image transmission method combining feature extraction and semantic compression, provided in an embodiment of the present invention;

[0049] Figure 2 This is a data flow diagram illustrating an underwater acoustic image transmission method combining feature extraction and semantic compression, provided by an embodiment of the present invention.

[0050] Figure 3 A schematic diagram illustrating the extraction of semantic feature sets provided in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of differential coding in an underwater acoustic image transmission method that combines feature extraction and semantic compression, provided by an embodiment of the present invention.

[0052] Figure 5 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0055] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0056] It should also be further understood that the term "and / or" as used herein, in the specification and in the claims, means any combination of one or more of the associated listed items and can be used interchangeably with "or". Moreover, it is also to be understood that, in the claims, "but not" does not necessarily indicate an exclusive disjunction.

[0057] As used in the specification and the appended claims, the term "if' can be construed to mean "when" or "once" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [a described condition or event] is detected" can be construed to mean "once it is determined" or "in response to determining" or "once [the described condition or event] is detected" or "in response to detecting [the described condition or event]" depending on the context.

[0058] Referring to Figure 1 in combination Figure 2 , the embodiment of the present application provides a kind of acoustic image transmission method combining feature extraction and semantic compression, by its unique system architecture and processing flow, can effectively adapt to the characteristics of underwater acoustic channel, realize image efficient and reliable transmission under limited bandwidth and high packet loss environment, and keep the availability of its semantic content, specifically, the method comprises the following steps:

[0059] S1, the original underwater acoustic image is extracted by the sending end semantic feature, and the semantic feature set containing target category information, target position information and scene semantics is obtained.

[0060] In specific implementation, the original underwater acoustic image is extracted by the sending end semantic feature, and the massive pixel data containing a large amount of redundant visual details are converted and condensed into the compact semantic feature set containing target category information, target position information and scene semantics, which fundamentally changes the paradigm of data transmission, from transmitting "everything seen" to transmitting "understood", that is, the information is intelligently purified at the source, so that the amount of data to be transmitted can be effectively reduced.

[0061] For example, referring to Figure 3 For a underwater acoustic image containing fish and seabed topography, this step does not transmit the brightness and color of each pixel point, but extracts key semantic elements such as "there are fish", "they are located in a specific area of the image" and "the background is sandy seabed". This greatly eliminates the redundant information irrelevant to the task, lays a solid foundation for subsequent data compression, and is the primary prerequisite for achieving efficient transmission.

[0062] In some preferred embodiments, the above step of "performing semantic feature extraction on the original underwater acoustic image to obtain a semantic feature set containing target category information, target position information and scene semantics" specifically comprises the following steps: performing preprocessing on the original underwater acoustic image to obtain a preprocessed image, the preprocessing including filtering and noise reduction and image enhancement; performing multi-level feature analysis on the preprocessed image using a pre-set deep learning network to extract multi-level features containing global feature vectors, local feature vectors, target category information and target position information; and fusing the extracted multi-level features to form the semantic feature set.

[0063] In specific implementation, first, by performing preprocessing including filtering and noise reduction and image enhancement on the original underwater acoustic image, the image quality degradation problem caused by the poor underwater imaging environment can be effectively overcome, such as reducing scattering noise and improving contrast, which enables the subsequent deep learning network to analyze on the basis of a clearer image with higher signal-to-noise ratio, avoiding that noise is misjudged as a feature, thereby improving the accuracy of feature extraction.

[0064] Further, the purpose of performing multi-level feature analysis using a pre-set deep learning network is to analyze the image content from different dimensions and granularities to ensure that comprehensive information from macro scene layout to micro target details is captured. The global feature vector carries the overall semantics of the scene, the local feature vector retains the detailed texture of the target, and the target category information and the target position information directly correspond to the task requirements.

[0065] Further, fusing the above multi-level features can form a unified and information complementary semantic feature set, which not only has a regular structure and is beneficial to subsequent encoding, but also contains information of different abstraction levels, so that the receiving end can obtain richer context when restoring, whether it is image reconstruction or task recognition, the fidelity and accuracy are effectively guaranteed.

[0066] In some preferred embodiments, the multi-level feature analysis includes: extracting a global feature vector representing scene-level background semantics and overall layout in the preprocessed image; extracting target category information and target position information of a key target in the preprocessed image, the target position information including a bounding box and / or a segmentation mask; and extracting a local feature vector representing texture, edge and shape inside a target region in the preprocessed image.

[0067] In specific implementation, the global feature vector representing scene-level background semantics and overall layout is extracted to enable the system to understand the macro environment in which the image is located, such as open water or complex seabed topography, which provides a crucial context for the interpretation of local targets and helps to improve the accuracy of target recognition and the depth of scene understanding.

[0068] Further, the target class information and target position information (including the bounding box and / or segmentation mask) of the extracted key target directly correspond to the most core task requirements in underwater monitoring, namely "what" and "where". The bounding box provides a rough location of the target, while the segmentation mask can accurately outline the contour of the target, which enables the receiving end to accurately locate and identify the target.

[0069] Further, local feature vectors representing texture, edge and shape are extracted within the target region. The purpose is to retain the fine appearance features of the target. Although the data volume of these features is relatively large, they are essential for distinguishing similar targets, conducting more detailed species identification, or restoring high-quality visual effects when reconstructing images. The three levels of features work together to form a complete semantic description from global to local and from rough to fine, ensuring that in subsequent processing, different levels of information can be flexibly selected without losing the core semantics according to the channel conditions and task requirements.

[0070] The structured representation of the final output semantic feature set is: S = {C, M, Z_g, Z_l};

[0071] Where Z_g is the global feature vector, Z_l is the local feature vector, C is the target class information, and M is the target position information

[0072] Specifically, in the present embodiment, the specific implementation of the above-mentioned step "using a pre-set deep learning network to perform multi-level feature analysis on the pre-processed image, extracting multi-level features including global feature vectors, local feature vectors, target class information and target position information; and fusing the extracted multi-level features to form the semantic feature set" is as follows:

[0073] First, a deep convolutional neural network containing an encoder-decoder structure is constructed. The encoder part uses a backbone network with a multi-level downsampling structure to extract feature maps of different scales from the pre-processed image. Among them, the deep feature map is transformed into a fixed-dimensional global feature vector after global average pooling operation, which is used to represent the background semantics and overall layout information at the scene level.

[0074] Further, a target detection branch is connected to the feature map output by the encoder. The target detection branch generates candidate regions through a region proposal network, extracts region features using a region of interest alignment layer, and then outputs target class information through a classifier and target bounding box coordinates through a regressor, thereby obtaining accurate target position information. Further, the image features of the candidate regions are extracted to obtain local feature vectors.

[0075] Further, the global feature vector, the local feature vector, the target category information, the bounding box coordinates and the segmentation mask are jointly input into a feature fusion module. The module calculates the correlation weight between each feature component based on an attention mechanism-based weighting fusion strategy, adaptively weights and combines the multi-level features, and finally forms a unified semantic feature set to provide a structured semantic representation for subsequent compression encoding.

[0076] In some preferred embodiments, before step S2, the following steps are further included: importance evaluation is performed on the features in the semantic feature set; based on the importance evaluation result, task-critical semantic features are screened and retained from the semantic feature set, and redundant features are discarded.

[0077] In a specific implementation, importance evaluation is performed on the features in the semantic feature set, and the purpose is to quantify the contribution of each feature to the completion of a specific underwater task (such as target detection, classification). For example, in a fish school monitoring task, the importance of the category and position information of fish is obviously higher than the details of the seabed sand texture. Through such evaluation, the system can identify which are task-critical semantic features and which are redundant features with lower contribution.

[0078] Further, based on the evaluation result, the critical features are retained and the redundant features are discarded, which is performed before compression encoding, equivalent to a second “purification” at the semantic level. This makes the data set finally entering the compression encoding link highly refined, containing only the most indispensable information for the successful execution of the task. Therefore, this step can significantly reduce the overall data volume of the compressed semantic data to be transmitted, avoid the precious underwater acoustic channel bandwidth being occupied by secondary or redundant information, and ensure that in the most tense bandwidth conditions, the most valuable data resources can be prioritized for transmitting information that can most guarantee the success of the task, greatly improving the efficiency and economy of transmission.

[0079] Specifically, in the present embodiment, the specific implementation of the above-mentioned step “importance evaluation is performed on the features in the semantic feature set; based on the importance evaluation result, task-critical semantic features are screened and retained from the semantic feature set, and redundant features are discarded” is as follows:

[0080] First, a task-driven feature importance evaluation model is constructed. The model receives the semantic feature set as input, calculates the gradient of each feature component with respect to the objective function of the preset underwater task through a multi-layer perception network, quantifies the contribution of the feature to the task output according to the gradient amplitude, and obtains a preliminary importance score.

[0081] Further, a redundancy analysis module based on mutual information is adopted to calculate the statistical dependency between different feature components. The mutual information value between features is measured by kernel density estimation method. When the mutual information between two features exceeds a preset threshold, it is determined that there is information redundancy, and the feature with lower relevance to the target task is down-weighted.

[0082] Further, an attention-weighted feature selection mechanism is introduced. The global feature vector, local feature vector, target class information and target position information are weighted and fused by a trainable attention weight matrix, and the feature dimensions that contribute significantly to the current transmission task are automatically identified and strengthened according to the attention weight.

[0083] Further, feature screening is implemented based on comprehensive importance scores. A dynamically adjustable importance threshold is set to retain task-critical semantic features with scores higher than the threshold, forming a reduced feature subset. The threshold can be dynamically adjusted according to real-time channel conditions, and a stricter screening standard is adopted when the bandwidth is limited to ensure that only the most core semantic information is retained.

[0084] S2, compressing and encoding the semantic feature set to generate compressed semantic data.

[0085] In specific implementation, the semantic feature set is compressed and encoded to generate compressed semantic data. Since the semantic feature set itself is a highly structured and abstract data representation, its inherent redundancy has been significantly reduced compared to the original pixel matrix. On this basis, targeted compression and encoding can further greatly reduce the data volume. This compression is not simply a size reduction, but an optimized encoding based on the data characteristics of semantic information, so as to generate a compressed semantic data packet with extremely small data volume while ensuring the integrity of the core semantics.

[0086] In some preferred embodiments, referring to Figure 4 The above step of "compressing and encoding the semantic feature set to generate compressed semantic data" specifically includes the following steps: hierarchically dictionary encoding the target class information; using differential encoding for the bounding box in the target position information, and using a hybrid compression method combining binarization and run-length encoding and block-based hash encoding for the segmentation mask in the target position information; using vector quantization and entropy encoding for the global feature vector; and using low-rank matrix approximation or clustering compression for the local feature vector.

[0087] In specific implementation, the target class information is hierarchically dictionary encoded. The principle is to assign short code words to high-frequency classes and long code words to low-frequency classes based on the discrete and frequency-differentiated characteristics of class labels, which conforms to the entropy encoding principle in information theory and can significantly reduce the average encoding length of class information.

[0088] Further, the bounding box in the target position information is differentially encoded, which is based on the characteristic that the position change between consecutive frames or adjacent targets is usually small, and only the position difference is encoded instead of the absolute coordinates, thereby effectively eliminating the space-time redundancy in the data.

[0089] Further, a hybrid compression method of binarization combined with run-length encoding and block-based hash encoding is used for the segmentation mask, the run-length encoding is good at compressing consecutive same pixels in the binary image, and the hash encoding can identify and merge repeated image blocks, and this combination strategy can efficiently handle the sparsity and repetition in the mask data.

[0090] Specifically, first, the segmentation mask is binarized to generate a binary image composed of foreground pixels and background pixels. Then, the binary image is run-length encoded, and the pixel points are sequentially traversed by scanning lines, and the consecutive same pixel value sequence is converted into a three-tuple data format of (start position, run length, pixel value), realizing efficient compression of sparse areas.

[0091] Further, for the dense area in the binary image, a block-based hash encoding method is used: the image is divided into fixed-size pixel blocks, and the hash fingerprint of each pixel block is calculated; a hash dictionary library is established to store the unique pixel block data, and only the hash index value and its position coordinates of each pixel block are stored at the encoding end.

[0092] Further, the sparse area data encoded by run-length encoding and the dense area data encoded by hash encoding are distinguished by metadata, and the compressed mask data stream is generated by combination. This hybrid compression method fully utilizes the spatial redundancy characteristics existing in the segmentation mask, and realizes significant data compression effect while ensuring the integrity of the key contour information.

[0093] Further, the global feature vector is quantized and entropy encoded, the vector quantization approximates the original high-dimensional vector with representative vectors in the codebook, realizing lossy but efficient compression, and then entropy encoding is used to further eliminate statistical redundancy.

[0094] Further, the local feature vector is compressed by low-rank matrix approximation or clustering, which utilizes the inherent structural redundancy in these feature tensors, and greatly reduces the data dimension by extracting the main components or representative centers. This kind of classified and customized encoding method can achieve higher compression ratio and better information retention than using a single general compression method for the original pixels or the mixed feature set.

[0095] In some preferred embodiments, the step of "compressively encoding the semantic feature set to generate compressed semantic data" further comprises the steps of: dividing the semantic feature set into a core layer and an enhancement layer, wherein the core layer comprises at least target class information, target location information and a global feature vector, and the enhancement layer comprises a local feature vector; and encoding and packing the core layer and the enhancement layer separately.

[0096] In specific implementations, the target class information, the target location information and the global feature vector are divided into the core layer because these information collectively form the minimum complete set for understanding the image content and performing basic tasks: knowing "what", "where" and "how the scene looks like" is enough to accomplish most detection, recognition and preliminary analysis tasks even without detailed textures.

[0097] Further, the local feature vector containing detailed textures is divided into the enhancement layer, which acknowledges that these information are mainly used to enhance user experience or conduct more refined analysis and can be sacrificed when communication resources are scarce. Encoding and packing the core layer and the enhancement layer separately makes the data clearly separated in structure. This division enables the sender and the receiver to treat these two parts of data differently, laying a foundation for imposing stronger error protection and assigning higher transmission priority on the core layer data.

[0098] S3, transmitting the compressed semantic data to the receiver through the underwater acoustic channel.

[0099] In specific implementations, the compressed semantic data is transmitted to the receiver through the underwater acoustic channel. Because the total amount of data to be transmitted is very small due to the aforementioned steps, the data packet can be effectively packed and transmitted even under harsh conditions where the bandwidth of the underwater acoustic channel is extremely limited, thereby significantly improving the probability and efficiency of successful transmission. The sharp reduction in data volume also means that the probability of successfully receiving the complete data packet is higher under the same error rate. At the same time, smaller data packets shorten the transmission time, which helps to reduce the end-to-end latency and reduce the risk of transmission interruption due to the time-varying characteristics of the channel.

[0100] In some preferred embodiments, the ratio of compressive encoding is dynamically adjusted according to the bandwidth condition of the current underwater acoustic channel, wherein only the compressed semantic data of the core layer is transmitted when the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold.

[0101] In specific implementations, the bandwidth of the underwater acoustic channel is not constant and will fluctuate dramatically due to various factors such as temperature, salinity, water depth and multipath effect. The purpose of dynamically adjusting the ratio of compressive encoding according to the bandwidth condition of the current underwater acoustic channel is to make the output data volume match the actual carrying capacity of the channel in real time.

[0102] Specifically, when the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold, only the compressed semantic data of the core layer is transmitted, which is an intelligent strategy for preserving key data. When the channel capacity is severely insufficient, the system actively gives up transmitting the local details for enhanced experience (enhancement layer) and instead concentrates all valuable bandwidth resources to ensure that the core layer information containing the target category, location, and scene semantics can be reliably delivered to the receiving end. This means that even in the case of extremely deteriorated channel conditions that cannot support complete image data transmission, the system can still ensure the transmission of the most core task-level information (such as "finding a suspicious target in a certain area"), thereby achieving uninterrupted service and significantly improving the survival ability and practical value of the system in complex underwater environments.

[0103] In some preferred embodiments, the above step of "transmitting the compressed semantic data to the receiving end through the underwater acoustic channel" specifically includes the following step: using an unequal error protection mechanism to channel encode the compressed semantic data, wherein the core layer data is allocated higher redundancy than the enhancement layer data.

[0104] In specific implementations, using the unequal error protection mechanism for channel encoding implements the data priority established in the previous step at the channel transmission level, significantly improving the transmission reliability of key data in poor channels through an asymmetric resource allocation strategy. The core principle is to recognize that the importance of core layer data and enhancement layer data to the final task completion of the system is not equal, and therefore they should not be provided with the same level of protection when combating channel errors and packet loss. Allocating higher redundancy to core layer data than to enhancement layer data means that more error correction check information or more redundant transmission symbols will be added to the core layer data during channel encoding. When data packets are transmitted in the underwater acoustic channel and encounter random or burst errors and packet loss, core layer data with higher redundancy has stronger self-repairing ability and fault tolerance. Even if some transmission symbols are lost, the receiving end still has a high probability of successfully decoding the complete core layer data using the remaining redundant information.

[0105] Further, for the enhancement layer data, since its importance is relatively low, it is allowed to bear a higher risk of loss. This differentiated protection strategy allows limited channel encoding resources (such as increased redundancy overhead) to be prioritized for protecting the most critical information, thereby improving the robustness and task completion rate of the system in high-pressure transmission environments with the highest efficiency overall.

[0106] In some preferred embodiments, the above step of "using an unequal error protection mechanism to channel encode the compressed semantic data" specifically includes the following steps: using an extended window fountain code to allocate a higher extraction probability to the data of the core layer when generating symbols, and using a progressive window expansion strategy, with the initial window covering the data of the core layer first.

[0107] In a specific implementation, the extended window fountain code is a rateless code, which is characterized by the ability to recover the original data from any sufficient number of encoded symbols. The higher extraction probability assigned to the data of the core layer when generating symbols means that the system is more inclined to select data from the core layer data for encoding when generating each transmission symbol. This results in the symbols representing the core layer information dominating in number in the final generated transmission symbol stream, i.e., the core layer data is implicitly given a higher redundancy.

[0108] Further, the progressive window extension strategy is adopted, and the initial window preferentially covers the data of the core layer, which further strengthens the priority. At the beginning of transmission, the encoder generates symbols based only on the core layer data, ensuring that the receiving end has a high probability of successfully decoding the core layer first when receiving the initial batch of symbols. As the transmission proceeds, the window is gradually extended to include the enhancement layer data. This mechanism enables the receiving end to achieve progressive decoding of "core first, enhancement later", i.e., even if the transmission is interrupted due to deteriorating channel conditions, as long as the core layer data has been successfully decoded, the basic task is ready to be completed, thereby greatly enhancing the system's tolerance to transmission interruption.

[0109] In some preferred embodiments, the channel coding adopts a joint source-channel coding strategy, in which: the core layer data is encoded using a digital fountain code; the enhancement layer data is encoded using a joint source-channel coding similar to simulation, embedding continuous features into the modulation symbol space.

[0110] In a specific implementation, the core layer data is encoded using a digital fountain code, which selects an absolutely reliable digital transmission path. Fountain codes can mathematically guarantee that the original data can be recovered without error as long as the number of received symbols reaches a threshold by increasing redundancy. This provides a solid "fortress" protection for the critical discrete information of the core layer (such as category labels, position coordinates), ensuring the accuracy of the task core instructions.

[0111] Further, the adoption of a joint source-channel coding method for the enhancement layer data is a more efficiency-oriented approach. This method directly maps the continuous local feature vectors to modulation symbols, skipping the traditional quantization and channel coding steps. The advantage is that each transmitted symbol carries meaningful information. In good channel conditions, the details can be recovered with high quality; when the channel has errors or packet loss, the recovered features will show a smooth degradation (e.g., blurred details), rather than a catastrophic, block-like decoding failure or complete information loss as in digital transmission. This hybrid "digital + analog" strategy ensures the absolute accuracy of the core instructions while allowing the enhancement information to smoothly transition in quality when bandwidth is limited, thus providing the best overall user experience under various channel conditions.

[0112] S4, the receiving end decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task recognition based on the semantic feature set.

[0113] In specific implementations, the receiving end decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task recognition based on the semantic feature set. This step is crucial to ensure the practicality and robustness of the entire system. The core of the receiving end is to understand and utilize semantics, rather than mechanically stitching pixels. Even if some data is lost during transmission, as long as the core semantic feature set can be successfully restored, the system has the ability to perform tasks.

[0114] For example, as long as the key semantic "fish and its position" is recovered, the receiving end can directly output a target detection report to achieve task recognition. If visualization is required, an approximate image with correct semantic content can be reconstructed based on the restored semantic feature set using a generative model. The reconstructed image is structurally and key-target-identifiable, rather than a pile of meaningless broken pixel blocks. This semantic-based recovery mechanism makes the system highly tolerant to errors and packet loss during transmission, ensuring that the output information remains highly usable in high-error and high-packet-loss underwater acoustic environments.

[0115] In some preferred embodiments, the step of "decoding the received compressed semantic data" specifically includes the following steps: decoding the data of the core layer first to restore the target class information, target position information, and global feature vector; and on the basis of successful decoding of the data of the core layer, gradually decoding the data of the enhancement layer to restore the local feature vector.

[0116] In implementation, the progressive decoding process of the receiving end makes the receiving end processing echo the priority transmission strategy of the sending end, realizing the fast and effective use of available information in uncertain transmission environment. The purpose of decoding the data of the core layer first is to start the decoding process as soon as the number of received transmission symbols containing core layer information reaches the decoding threshold, to recover the target category information, target location information and global feature vector. This means that the receiving end can obtain all the information needed to perform the key task without waiting for all the data (including the enhancement layer) to be completely and correctly received. This "receiving and processing" mode greatly reduces the system response delay, and is especially suitable for underwater monitoring scenarios with high real-time requirements.

[0117] Further, on the basis of successful decoding of the data of the core layer, the data of the enhancement layer is continuously decoded in a progressive manner, which embodies a resource on-demand loading strategy. With the arrival of more transmission symbols (which may begin to contain more enhancement layer information), the receiving end continues to attempt to recover the local feature vector. This progressive feature enables the reconstruction quality of the image or the fineness of the analysis to gradually improve as the data transmission progresses. The entire process ensures that the system can provide services based on the highest priority information that has been successfully decoded at any time, realizes a smooth transition from "available" to "good use", and maximizes the utilization rate of the results of imperfect transmission process.

[0118] In some preferred embodiments, the above step "image reconstruction and / or task recognition based on the semantic feature set" specifically includes the following steps: decoding the data of the core layer first to recover the target category information, target location information and global feature vector; on the basis of successful decoding of the data of the core layer, continue to decode the data of the enhancement layer in a progressive manner to recover the local feature vector.

[0119] In implementation, when only the data of the core layer is successfully decoded, the recovered target category information, target location information and global feature vector are directly used to output the task recognition result, which is a very practical degradation strategy. It acknowledges that in the case of extremely scarce communication resources, perfect recovery of visual images may be a luxury, but completing the task itself is the ultimate goal. For example, even if a clear image cannot be seen, the system can output a structured report such as "Region A discovers fish, quantity about X", which is already fully sufficient for many automated monitoring tasks.

[0120] Further, when the data of the core layer and the enhancement layer is successfully decoded, the original underwater image is reconstructed based on the complete semantic feature set using the generation model, and the task recognition result is output, thereby providing a functional and high-quality service. At this time, the system can not only provide an accurate task report, but also provide visual image feedback for the operator, thereby facilitating manual interpretation, archiving, or more complex analysis. The ability to adaptively select the output mode according to the actual receiving condition enables the transmission method to have flexibility from 'only guaranteeing the task' to 'balancing the task and experience', thereby enabling the method to always provide maximized available output in a severely changing underwater acoustic communication environment, and greatly improving the practicality and user satisfaction of the method.

[0121] The embodiment of the present application provides an underwater acoustic image transmission method combining feature extraction and semantic compression, which comprises: performing semantic feature extraction on an original underwater acoustic image at a sending end to obtain a semantic feature set containing target category information, target position information and scene semantics; performing compression encoding on the semantic feature set to generate compressed semantic data; transmitting the compressed semantic data to a receiving end through an underwater acoustic channel; and performing decoding on the received compressed semantic data at the receiving end, restoring the semantic feature set, and performing image reconstruction and / or task recognition based on the semantic feature set. The present application converts pixel data into a semantic feature set containing target category, position and scene semantics by performing semantic feature extraction on an underwater acoustic image at a sending end; then performs compression encoding to significantly reduce the data volume, so that the data volume is adapted to the limited bandwidth of the underwater acoustic channel; and performs image reconstruction or direct task recognition based on the restored semantic features at a receiving end. The method realizes efficient and reliable transmission of image data in a low-bandwidth and high-packet-loss underwater acoustic environment, and ensures effective maintenance of semantic information and task availability.

[0122] Please refer to Figure 5 , Figure 5 is a schematic block diagram of a computer device provided by the embodiment of the present application. The computer device 500 can be a terminal or a server, wherein the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a personal digital assistant, a wearable device, and the like. The server can be a stand-alone server or a server cluster composed of multiple servers.

[0123] The computer device 500 comprises a processor 502, a memory and a network interface 505 connected through a system bus 501, wherein the memory can comprise a non-volatile storage medium 503 and an internal memory 504.

[0124] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, the processor 502 can execute an underwater acoustic image transmission method combining feature extraction and semantic compression.

[0125] The processor 502 is configured to provide computing and control capabilities to support the operation of the entire computer device 500.

[0126] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503, which, when executed by the processor 502, causes the processor 502 to perform a method of underwater acoustic image transmission combining feature extraction and semantic compression.

[0127] The network interface 505 is configured to communicate with other devices via a network. Those skilled in the art can understand that the above structure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device 500 to which the scheme of the present application is applied. The specific computer device 500 can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0128] The processor 502 is configured to run the computer program 5032 stored in the memory to implement the following steps:

[0129] The sender extracts semantic features from the original underwater acoustic image to obtain a semantic feature set containing target category information, target position information, and scene semantics;

[0130] The semantic feature set is compressed and encoded to generate compressed semantic data;

[0131] The compressed semantic data is transmitted to the receiver through the underwater acoustic channel;

[0132] The receiver decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task recognition based on the semantic feature set.

[0133] Further technical solutions are as follows: the semantic feature extraction from the original underwater acoustic image to obtain a semantic feature set containing target category information, target position information, and scene semantics comprises:

[0134] The original underwater acoustic image is preprocessed to obtain a preprocessed image, and the preprocessing includes filtering and denoising and image enhancement;

[0135] A pre-set deep learning network is used to analyze the preprocessed image at multiple levels to extract multi-level features including global feature vectors, local feature vectors, target category information, and target position information;

[0136] The extracted multi-level features are fused to form the semantic feature set.

[0137] Further technical solutions are as follows: the multi-level feature analysis comprises:

[0138] extracting a global feature vector representing scene-level background semantics and overall layout in the pre-processed image;

[0139] extracting target class information and target location information of key targets in the pre-processed image, the target location information including a bounding box and / or a segmentation mask;

[0140] extracting a local feature vector representing texture, edge and shape within a target region in the pre-processed image.

[0141] Further, the method further comprises, before the semantic feature set is compressed and encoded to generate compressed semantic data:

[0142] performing importance evaluation on the features in the semantic feature set;

[0143] based on the importance evaluation result, filtering and retaining task-critical semantic features from the semantic feature set, and discarding redundant features.

[0144] Further, the compressed and encoded semantic feature set is generated by:

[0145] adopting hierarchical dictionary encoding for the target class information;

[0146] adopting differential encoding for the bounding box in the target location information, and adopting a hybrid compression method combining binarization and run-length encoding and block-based hash encoding for the segmentation mask in the target location information;

[0147] adopting vector quantization and entropy encoding for the global feature vector;

[0148] adopting low-rank matrix approximation or clustering compression for the local feature vector.

[0149] Further, the compressed and encoded semantic feature set is generated by:

[0150] dividing the semantic feature set into a core layer and an enhancement layer, wherein the core layer at least includes the target class information, the target location information and the global feature vector, and the enhancement layer includes the local feature vector;

[0151] encoding and packing the core layer and the enhancement layer respectively;

[0152] wherein, according to the bandwidth condition of the current underwater acoustic channel, the compression and encoding ratio is dynamically adjusted, and when the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold, only the compressed semantic data of the core layer is transmitted.

[0153] Further, the technical scheme is that the compressed semantic data is transmitted to a receiving end through a water acoustic channel, and the method comprises the following steps:

[0154] The compressed semantic data is channel encoded by using an unequal error protection mechanism, wherein higher redundancy is allocated to core layer data than to enhancement layer data.

[0155] The channel encoding of the compressed semantic data comprises the following steps:

[0156] An extended window fountain code is used to allocate higher extraction probability to the core layer data when the symbol is generated, and a progressive window extension strategy is used, and the initial window preferentially covers the core layer data.

[0157] Further, the technical scheme is that the channel encoding adopts a joint source channel coding strategy, wherein:

[0158] The core layer data is encoded by using a digital fountain code;

[0159] The enhancement layer data is encoded by using a joint source channel coding similar to simulation, and continuous features are embedded into a modulation symbol space.

[0160] Further, the technical scheme is that the received compressed semantic data is decoded, and the method comprises the following steps:

[0161] The core layer data is preferentially decoded to restore target category information, target position information and a global feature vector;

[0162] On the basis of successful decoding of the core layer data, the enhancement layer data is progressively decoded to restore a local feature vector.

[0163] Further, the technical scheme is that the image is reconstructed and / or task recognition is performed based on the semantic feature set, and the method comprises the following steps:

[0164] When only the core layer data is successfully decoded, the restored target category information, target position information and global feature vector are directly used to output a task recognition result;

[0165] When the core layer data and the enhancement layer data are successfully decoded, the original water acoustic image is reconstructed based on the complete semantic feature set by using a generative model, and a task recognition result is output.

[0166] It should be understood that, in the embodiments of the present application, the processor 502 can be a central processing unit (CPU), and the processor 502 can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0167] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to realize the process steps of the above-mentioned embodiments.

[0168] Therefore, the present application also provides a storage medium. The storage medium can be a computer-readable storage medium. The storage medium stores a computer program. The computer program is executed by a processor to make the processor perform the following steps:

[0169] The sender extracts semantic features from the original underwater acoustic image to obtain a semantic feature set containing target category information, target position information and scene semantics;

[0170] The semantic feature set is compressed and encoded to generate compressed semantic data;

[0171] The compressed semantic data is transmitted to the receiver through an underwater acoustic channel;

[0172] The receiver decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task recognition based on the semantic feature set.

[0173] Further technical solutions are that the semantic feature extraction from the original underwater acoustic image to obtain a semantic feature set containing target category information, target position information and scene semantics comprises:

[0174] The original underwater acoustic image is preprocessed to obtain a preprocessed image, and the preprocessing includes filtering and denoising and image enhancement;

[0175] performing multi-level feature analysis on the preprocessed image by using a preset deep learning network to extract multi-level features including global feature vectors, local feature vectors, target category information and target position information;

[0176] fusing the extracted multi-level features to form the semantic feature set.

[0177] A further technical solution is that the multi-level feature analysis includes:

[0178] extracting global feature vectors representing scene-level background semantics and overall layout in the preprocessed image;

[0179] extracting target category information and target position information of key targets in the preprocessed image, the target position information including a bounding box and / or a segmentation mask;

[0180] extracting local feature vectors representing texture, edge and shape inside a target region in the preprocessed image.

[0181] A further technical solution is that before the semantic feature set is compressed and encoded to generate compressed semantic data, the method further includes:

[0182] performing importance evaluation on features in the semantic feature set;

[0183] based on the importance evaluation result, filtering and retaining task-critical semantic features from the semantic feature set and discarding redundant features.

[0184] A further technical solution is that the compression and encoding of the semantic feature set to generate compressed semantic data includes:

[0185] adopting hierarchical dictionary encoding for the target category information;

[0186] adopting differential encoding for the bounding box in the target position information and adopting a hybrid compression method combining binarization and run-length encoding and block-based hash encoding for the segmentation mask in the target position information;

[0187] adopting vector quantization and entropy encoding for the global feature vectors;

[0188] adopting low-rank matrix approximation or clustering compression for the local feature vectors.

[0189] A further technical solution is that the compression and encoding of the semantic feature set to generate compressed semantic data further includes:

[0190] The semantic feature set is divided into a core layer and an enhancement layer, wherein the core layer at least includes target category information, target position information and a global feature vector, and the enhancement layer includes a local feature vector;

[0191] The core layer and the enhancement layer are respectively encoded and packaged;

[0192] The ratio of compression encoding is dynamically adjusted according to the bandwidth condition of the current underwater acoustic channel, wherein when the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold, only the compressed semantic data of the core layer is transmitted.

[0193] A further technical solution is that the compressed semantic data is transmitted to a receiving end through an underwater acoustic channel, comprising:

[0194] An unequal error protection mechanism is used for channel encoding of the compressed semantic data, wherein the core layer data is allocated higher redundancy than the enhancement layer data;

[0195] The channel encoding of the compressed semantic data using the unequal error protection mechanism comprises:

[0196] An extended window fountain code is used to allocate higher extraction probability to the data of the core layer when the symbol is generated, and a progressive window extension strategy is used, and the initial window preferentially covers the data of the core layer.

[0197] A further technical solution is that the channel encoding uses a joint source channel encoding strategy, wherein:

[0198] The core layer data is encoded using a digital fountain code;

[0199] The enhancement layer data is encoded using a similar analog joint source channel encoding, and the continuous features are embedded into a modulation symbol space.

[0200] A further technical solution is that the received compressed semantic data is decoded, comprising:

[0201] The data of the core layer is preferentially decoded to restore the target category information, the target position information and the global feature vector;

[0202] On the basis of successful decoding of the data of the core layer, the data of the enhancement layer is continuously decoded in a progressive manner to restore the local feature vector.

[0203] A further technical solution is that the image reconstruction and / or task recognition are performed based on the semantic feature set, comprising:

[0204] When only the data of the core layer is successfully decoded, the restored target category information, target position information and global feature vector are directly used to output the task recognition result;

[0205] When the data of the core layer and the enhancement layer are successfully decoded, the original underwater image is reconstructed based on the complete semantic feature set by using the generation model, and a task recognition result is output.

[0206] The storage medium is a physical, non-transient storage medium, for example, can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, and various physical storage media that can store program codes. The computer-readable storage medium can be non-volatile or volatile.

[0207] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in the above description in general terms. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0208] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0209] The steps in the method embodiments of the present application can be adjusted, combined and reduced in sequence according to actual needs. The units in the device embodiments of the present application can be combined, divided and reduced according to actual needs. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0210] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a terminal or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application.

[0211] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0212] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0213] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for underwater acoustic image transmission combining feature extraction and semantic compression, characterized in that, include: The transmitting end extracts semantic features from the original underwater acoustic image to obtain a semantic feature set containing target category information, target location information, and scene semantics; The semantic feature set is compressed and encoded to generate compressed semantic data; The compressed semantic data is transmitted to the receiving end via an underwater acoustic channel; The receiving end decodes the received compressed semantic data, restores the semantic feature set, and performs image reconstruction and / or task recognition based on the semantic feature set.

2. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 1, characterized in that, The semantic feature extraction of the original underwater acoustic image yields a semantic feature set containing target category information, target location information, and scene semantics, including: The original underwater acoustic image is preprocessed to obtain a preprocessed image, the preprocessing including filtering and noise reduction and image enhancement; The preprocessed image is subjected to multi-level feature analysis using a pre-defined deep learning network to extract multi-level features including global feature vectors, local feature vectors, target category information, and target location information. The extracted multi-level features are fused to form the semantic feature set.

3. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 2, characterized in that, The multi-level feature analysis includes: Extract global feature vectors representing scene-level background semantics and overall layout from the preprocessed image; Extract target category information and target location information of key targets from the preprocessed image, wherein the target location information includes bounding boxes and / or segmentation masks; Local feature vectors representing texture, edges, and shape are extracted within the target region of the preprocessed image.

4. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 1, characterized in that, Before compressing and encoding the semantic feature set to generate compressed semantic data, the method further includes: The importance of the features in the semantic feature set is evaluated; Based on the importance assessment results, key semantic features of the task are selected and retained from the set of semantic features, while redundant features are discarded.

5. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 2, characterized in that, The step of compressing and encoding the semantic feature set to generate compressed semantic data includes: The target category information is encoded using hierarchical dictionary encoding; Differential coding is used for the bounding box in the target location information, and a hybrid compression method combining binarization, run-length encoding, and block-based hash encoding is used for the segmentation mask in the target location information. Vector quantization and entropy encoding are applied to the global feature vector; Low-rank matrix approximation or cluster compression is used for local eigenvectors.

6. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 5, characterized in that, The step of compressing and encoding the semantic feature set to generate compressed semantic data further includes: The semantic feature set is divided into a core layer and an enhancement layer, wherein the core layer includes at least target category information, target location information and global feature vector, and the enhancement layer includes local feature vector; The core layer and the enhancement layer are encoded and packaged separately. The compression coding ratio is dynamically adjusted according to the current bandwidth conditions of the underwater acoustic channel. When the bandwidth of the current underwater acoustic channel is lower than a preset bandwidth threshold, only the compressed semantic data of the core layer is transmitted.

7. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 6, characterized in that, The step of transmitting the compressed semantic data to the receiving end via an underwater acoustic channel includes: An unequal error protection mechanism is used to perform channel coding on compressed semantic data, in which higher redundancy is allocated to the core layer data than to the enhancement layer data. The method of using an unequal error protection mechanism to perform channel coding on compressed semantic data includes: An extended window fountain code is adopted, which assigns a higher extraction probability to the core layer data during symbol generation, and a progressive window expansion strategy is adopted, with the initial window prioritizing the coverage of the core layer data.

8. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 7, characterized in that, The channel coding employs a joint source channel coding strategy, wherein: The core layer data is encoded using digital fountain codes; The enhancement layer data is encoded using a simulated joint source channel coding method, embedding continuous features into the modulation symbol space.

9. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 8, characterized in that, Decoding the received compressed semantic data includes: Prioritize decoding the core layer data to recover target category information, target location information, and global feature vectors; Based on the successful decoding of the core layer data, the data of the enhancement layer is then progressively decoded to recover the local feature vectors.

10. The underwater acoustic image transmission method combining feature extraction and semantic compression according to claim 9, characterized in that, The image reconstruction and / or task recognition based on the semantic feature set includes: When only the core layer data is successfully decoded, the task recognition result is directly output using the restored target category information, target location information, and global feature vector. When the data of the core layer and enhancement layer are successfully decoded, the original underwater acoustic image is reconstructed based on the complete semantic feature set using the generative model, and the task recognition result is output.

Images