Deep learning-based extraterrestrial planet life mark identification method and system

Abstract

The invention provides an extraterrestrial planet life trace identification method and system based on deep learning, relates to the field of planet remote sensing, ancient biology, stratigraphy and celestial body biology research, and aims to solve the problems that in the prior art, no comprehensive and effective ancient biological picture database exists, parameter data and model training data are lacked, and the life trace identification accuracy is poor. And the model identification capability is difficult to meet the identification requirements of life marks in extraterrestrial planet return samples. According to the method, a paleontology fossil picture database is constructed, paleontology fossil pictures are preprocessed, and a constructed paleontology vital sign recognition model is trained based on the preprocessed pictures; and carrying out vital sign identification on the to-be-identified paleontological fragment picture by using the trained paleontological vital sign identification model. The method solves the problem of lack of extraterrestrial planet life mark reference data, improves the recognition efficiency and recognition accuracy of an ancient organism vital sign recognition model, and enables the model to meet the requirements of future extraterrestrial planet life mark recognition.

Description

Technical Field

[0001] The present invention belongs to the research fields of planetary remote sensing, paleontology and stratigraphy, and astrobiology, and in particular relates to a method and system for identifying traces of life on extraterrestrial planets based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the advancement of deep space exploration technology, the search for traces of life is a key scientific goal for future deep space exploration missions, including exploration of exoplanets such as Jupiter's moons, Saturn, and Venus. This search for traces of life is a current focus of solar system exploration and planetary science. It will help answer the crucial scientific question of "Is humanity alone in the universe?" and has significant scientific significance for the study of the origin and evolution of life. Therefore, upon the return of samples from Mars and other extraterrestrial bodies, accurately identifying individual life units, fragments, or traces of extraterrestrial life through various analytical methods will be a crucial and challenging task.

[0004] Currently, existing technologies are typically based on laboratory visual interpretation or machine analysis, comparing Earth's paleontological fossils to identify traces of life on Mars and other exoplanets in sample microscopic images. However, the following problems still exist:

[0005] (1) The relevant knowledge is vast. According to previous studies, there are many kinds of paleontology on Earth (estimated to be 200,000 species). Training a professional paleontology fossil researcher requires huge manpower and material resources, and it is also difficult to find an expert or system that can master and understand the characteristics of all paleontology fossils on Earth.

[0006] (2) Poor application generalization. More than 99.9% of the organisms that once lived on Earth are now extinct. Moreover, the traces of life that may exist on Mars and other exoplanets may not be completely similar to Earth life in terms of morphology and evolution. Therefore, using existing Earth fossils to directly compare suspected life structures in returned samples will have a relatively poor recognition effect.

[0007] (3) Fragmentation of life traces. The possible time range for the emergence of Martian life is the Noachian period (about 4.1-3.7 billion years ago, a warm and humid environment). Subsequently, the environment and climate changed, and it experienced long-term and multiple weathering processes. Therefore, even if traces of ancient life once existed and are preserved to this day, the paleontological fossils in the returned samples may be severely damaged and fragmented, making their identification more difficult, and thus low in identification efficiency and accuracy.

[0008] It can be seen that the existing technology for identifying traces of life on extraterrestrial planets does not have a comprehensive and correct database of paleontological fragments that can guide the search for traces of life on Mars and extraterrestrial planets, and lacks a reference basis for paleontological life characteristics; at the same time, the existing recognition model is difficult to directly apply to the identification of traces of life in samples returned from extraterrestrial planets, and the generalization ability of the model recognition cannot meet the requirements. Summary of the Invention

[0009] To overcome the shortcomings of the aforementioned prior art, the present invention provides a deep learning-based method and system for identifying signs of life on extraterrestrial planets. This system constructs a database of paleontological fossil images and performs preprocessing, such as fragmentation. A model for identifying paleontological life signatures is then trained. The trained model is then used to identify microscopic images of samples returned from extraterrestrial bodies, thereby identifying signs of paleontological life. This method addresses the lack of reference data for extraterrestrial life signatures and improves the model's ability to identify extraterrestrial life signatures.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention discloses a method for identifying life traces on extraterrestrial planets based on deep learning, comprising:

[0012] Constructing a paleontological fossil image database and preprocessing the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-category multi-body images;

[0013] Based on the pre-processed images, the constructed paleontological life feature recognition model is trained;

[0014] Using the trained paleontological life feature recognition model, the paleontological fragment images to be identified are used to identify life features and obtain recognition results;

[0015] Among them, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is: extracting the first feature map of the preprocessed image through the basic network module; inputting the feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; based on the fused feature map, generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map; and performing target classification and position prediction on the image in the anchor frame through the category prediction and position prediction blocks.

[0016] As a further technical solution, a paleontological fossil image database is constructed. Specifically, according to the collection rules, images of Earth paleontological fossils from different eras related to extraterrestrial planets are collected, and they are labeled and accurately classified to obtain a paleontological fossil image database including images of protozoa, lower animals and plant pollen fossils.

[0017] As a further technical solution, the collection rules include:

[0018] Select molded fossils and chemical fossils as fossil types, and collect fossils of ancient Earth organisms observed under a microscope at millimeter to micrometer levels;

[0019] Collecting spores and pollen fossils of protists, lower animals and plants on Earth;

[0020] The collected fossils of reef-building organisms include individuals and complexes.

[0021] As a further technical solution, pre-process the paleontological fossil images. The specific process is as follows:

[0022] Fragmenting the single image to obtain a fragmented image;

[0023] Classifying and labeling the single image and the fragmented image;

[0024] Classify and label multi-class and multi-body images.

[0025] As a further technical solution, fragmentation processing is performed on the individual images, which refers to random rectangular cropping of the individual images of paleontological fossils. The specific process is as follows:

[0026] Setting a coordinate system with the lower left corner of the single image as the origin and determining the image size parameters;

[0027] Based on the coordinate system and image size parameters, a cropping rectangle parameter set is randomly generated, including the center point coordinates, area percentage, and aspect ratio;

[0028] Calculating the size and boundary coordinates of the cropping rectangle according to the cropping rectangle parameter group;

[0029] Determine whether the cropping rectangle is completely within the original single image range. If so, store the cropping rectangle parameter group in the valid parameter group set. If not, regenerate the cropping matrix parameter group.

[0030] Comparing the clipping matrix parameter group with the parameter groups in the valid parameter group set, and regenerating them if they are the same, until a predetermined number of unique parameter groups are obtained;

[0031] Batch cropping the original single image according to a predetermined number of unique parameter groups to generate a predetermined number of fragmented images;

[0032] Repeat the above steps to obtain a predetermined number of fragmented images of all the single images in sequence.

[0033] As a further technical solution, the first feature map is input into a multi-scale feature block, and feature enhancement is performed through an attention module and a scale transformation block. The multi-scale feature block includes multiple groups of attention modules and scale transformation blocks alternately connected in series. The specific process is as follows:

[0034] Inputting the first feature map into a first attention module for spatial feature enhancement to obtain a first enhanced feature map;

[0035] Performing a scale transformation on the first enhanced feature map by a scale transformation block to obtain a first scale feature map;

[0036] Inputting the first scale feature map into the second attention module for spatial feature enhancement to obtain a second enhanced feature map;

[0037] Performing a scale transformation on the second enhanced feature map by a scale transformation block to obtain a second scale feature map;

[0038] Based on the above steps, the input is iterated to each group of attention modules and scale transformations in turn until feature maps of all scales are obtained.

[0039] As a further technical solution, the first feature map is input into the first attention module for spatial feature enhancement. The specific process is as follows:

[0040] Performing one-dimensional global average pooling on the first feature map along the height and width respectively to generate height-direction features and width-direction features;

[0041] After concatenating the height-direction features and the width-direction features, the dimensionality is reduced through convolution to obtain the intermediate features.

[0042] Decompose and transpose the intermediate features to obtain width features for restoring width and height features for restoring height respectively;

[0043] The width feature of the restored width and the height feature of the restored height are convolved and dimensioned respectively, and the width and height attention weights are generated through the Sigmoid activation function;

[0044] The width and height attention weights are multiplied with the multi-scale feature map to obtain enhanced target features.

[0045] As a further technical solution, the multi-scale feature maps are fused through a feature fusion module, wherein the feature fusion module adopts a bidirectional pyramid network, including a bidirectional cross-scale connection structure, a depth-wise separable convolution layer, a weighted fusion mechanism module and a depth-wise separable convolution layer. The specific process is as follows:

[0046] In the bidirectional pyramid network, bidirectional feature propagation from top to bottom and from bottom to top is performed through a bidirectional cross-scale connection structure;

[0047] The multi-scale feature maps are weightedly fused through the weighted fusion mechanism module to obtain different scale feature map outputs;

[0048] Each scale feature output is processed through a depth-wise separable convolutional layer to obtain a processed multi-scale feature map;

[0049] The processed multi-scale feature map is downsampled or upsampled through the Resize operation to obtain the aligned multi-scale feature map;

[0050] The normalized fusion method is used to fuse the aligned scale feature maps to obtain the fused feature maps.

[0051] The second aspect discloses a deep learning-based system for identifying life traces on extraterrestrial planets, including:

[0052] A database construction and data preprocessing module is used to construct a paleontological fossil image database and preprocess the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-class multi-body images;

[0053] A model training module is used to train a constructed paleontological life feature recognition model based on preprocessed images; wherein, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is as follows: extracting the first feature map of the preprocessed image through the basic network module; inputting the first feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; based on the fused feature map, generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map; and performing target classification and position prediction on the image within the anchor frame through the category prediction and position prediction blocks;

[0054] The model recognition module is used to use the trained paleontological life feature recognition model to perform life feature recognition on the paleontological fragment images to obtain recognition results.

[0055] The third aspect of the present invention discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the program.

[0056] One or more of the above technical solutions have the following beneficial effects:

[0057] In this embodiment, a database of images of paleontological fossils on Earth is constructed, and the fossil images are fragmented and cropped. This can provide a comprehensive and accurate database foundation for identifying life characteristics and traces in future samples returned from exoplanets. This not only solves the problem of lack of reference data on traces of life on exoplanets, but also improves the recognition efficiency and accuracy of the paleontological life characteristic recognition model, making the model meet the requirements for future identification of traces of life on exoplanets.

[0058] In this embodiment, a paleontological life signature recognition model is constructed to extract life signatures from paleontological fossils of different types and scales, improving the ability to identify fragmented samples. Compared to existing models, this model does not directly and comprehensively compare paleontological fossils to search for extraterrestrial life and identify life signatures in samples returned from extraterrestrial planets. Instead, based on Earth's paleontological fossil database, the model extracts and classifies life signatures through feature extraction, feature fusion, and the coordinated cooperation of various modules such as category prediction and position prediction. Furthermore, an attention mechanism and optimized loss function are used to address model bias caused by imbalanced sample categories. The paleontological life signature recognition model in this embodiment can accurately identify paleontological life signatures in microscopic images and can be used in the future to search for life signatures in samples returned from extraterrestrial planets, providing a powerful auxiliary tool for deep space exploration.

[0059] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0061] Figure 1 This is a diagram showing the architecture of a paleontological life feature recognition model for a method for identifying traces of life on extraterrestrial planets based on deep learning, according to the first embodiment of the present invention;

[0062] Figure 2 Schematic diagram of the weighted fusion mechanism of BiFPN according to the first embodiment of the present invention;

[0063] Figure 3 A single paleontological fossil image in the paleontological fossil image database of the first embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the fragmentation of paleontological monomers according to Example 1 of the present invention;

[0065] Figure 5 This is the fragment identification result in the paleontological database test set of Example 1 of the present invention;

[0066] Figure 6 This is the out-of-database paleontological fossil image recognition result of Example 1 of the present invention. DETAILED DESCRIPTION

[0067] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0068] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0069] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0070] Example 1

[0071] This embodiment discloses a method for identifying traces of life on extraterrestrial planets based on deep learning.

[0072] To more clearly illustrate this embodiment, a deep learning-based process for identifying life traces on extraterrestrial planets can be specifically described as follows:

[0073] This paper provides a deep learning-based method for identifying life traces on extraterrestrial planets, including:

[0074] S1. Constructing a paleontological fossil image database and preprocessing the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-class multi-body images;

[0075] S2. Based on the preprocessed image, the constructed paleontological life feature recognition model is trained; wherein, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is: extracting the first feature map of the preprocessed image through the basic network module; inputting the first feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; based on the fused feature map, generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map; performing target classification and position prediction on the image within the anchor frame through the category prediction and position prediction blocks;

[0076] S3. Use the trained paleontological life feature recognition model to perform life feature recognition on the paleontological fragment images to obtain recognition results.

[0077] like Figure 1 As shown, in step S1, a paleontological fossil image database is constructed and the paleontological fossil images are preprocessed, wherein the image formats in the paleontological fossil image database are: single-body images and multi-category multi-body images.

[0078] S1-1. Build a paleontological fossil image database.

[0079] According to the collection rules, images of paleontological fossils from different eras are collected, labeled and accurately classified to obtain a paleontological fossil image database including images of protozoa, lower animals and plant pollen fossils.

[0080] Specifically, although the Earth is about 4.6 billion years old, biological fossils have only been found in strata dating back 3.5 billion years. The geological age of Mars is divided into four stages: the Pre-Noachian (about 4.6-4.1 billion years ago), the Noachian (about 4.1-3.7 billion years ago), the Occidental (about 3.7-3 billion years ago), and the Amazonian (3 billion years ago to present). The Noachian may have been a warm and humid environment, and the Occidental environment changed from warm and humid to dry and cold. Therefore, ancient life may have been produced in the Noachian, and its individuals, debris, and traces of life may have been preserved in subsequent environments. Since Martian life has undergone long periods of time and multiple weathering processes (space weathering, water and wind erosion, solar radiation, etc.) since its emergence, traces of life in samples returned from Mars and other extraterrestrial planets may be fossil fragments of lower life. Therefore, when collecting images of Earth's paleontological fossils to build a database, the following collection rules should be followed:

[0081] (1) For the type of fossils, choose mold fossils or chemical fossils; for the size of fossils, choose microfossils.

[0082] Due to technical and cost limitations, the amount of samples returned from Mars (and other exoplanets) is small, shallow, and primarily composed of surface soil. Therefore, there is a high probability that no solid fossils or trace fossils of ancient organisms will be found. Furthermore, after long periods of weathering, any traces of life in the samples are likely to exist in the form of fragments (microfossils). Therefore, the primary focus is on collecting microscopically observable terrestrial fossils at the millimeter to micrometer scale.

[0083] (2) Focus on lower organisms, while taking higher organisms into consideration.

[0084] Currently, most speculation about Martian life focuses on lower life forms, such as protozoa and lower plants and animals. However, the possibility that higher life forms once existed on Mars cannot be completely ruled out. Furthermore, the present invention does not identify (identical) Martian life by comparing it with individual Earth life forms, but rather by comparing shared characteristics between Martian and Earth life forms (morphology, texture, and structure, etc.). Therefore, in addition to collecting fossils of lower life forms on Earth, the database should also include a certain number of lower animal and plant fossils to extract their unique life characteristics.

[0085] (3) Reef-building organisms include individuals and complexes.

[0086] Some paleontological fossils are formed by lower life forms (such as individual algae) living in colonies, capturing, binding, and cementing sedimentary particles to form complex chromatographic structures. For these colonial reef-building organisms, the morphology and characteristics of individual and complex fossils differ. To fully capture the characteristics of life on Earth, fossils formed by these reef-building organisms should be recorded as two categories in the database, recording both the individual and overall growth layer structures.

[0087] According to the above collection rules, the paleontological fossil image database mainly includes the following types of earth paleontological fossil individual images: bacteria, algae, stromatolite individuals and complexes, stromatolite individuals and complexes, corals, ostracods, radiolarians, foraminifera, Conodonts, small shells, chitin insects, plant spores, etc.

[0088] To extract the life characteristics of Earth's paleontological fossils, we must first establish an Earth's paleontological fossil image database. Based on the principle of mainly collecting microfossils of lower life (protists, lower animals and plant pollen, etc.), we have currently collected more than 20 types of paleontological fossils, with a total of more than 25,000 images. We will continue to enrich the types of fossils and the total number of images in the database. Figure 3 Shown are photos of individual paleontological fossils of typical types in the database.

[0089] After the above steps, a database of Earth paleontological images will be established to solve the problem that there is currently no comprehensive and correct database of paleontological fragments images that can guide the search for traces of life on Mars and extraterrestrial planets, and to provide a comprehensive and effective data foundation for subsequent operations and model identification.

[0090] S1-2. Preprocess the paleontological fossil images.

[0091] The images in the constructed paleontological fossil image database include two forms: one is a single image containing only one type of paleontology, and the other is an image containing multiple species and multiple paleontological individuals, which also includes complete microscopic images of the matrix and rock fragments.

[0092] (1) Fragment the single image to obtain fragmented images, and classify and label the single image and fragmented images.

[0093] Fragmenting a single image means randomly cropping a single image of a paleontological fossil into a rectangular shape. The specific process is as follows:

[0094] 1) Set a coordinate system with the lower left corner of the single image as the origin and determine the image size parameters.

[0095] In this embodiment, the lower left corner of the image to be cropped is set as the origin, the y-axis is upward and the x-axis is right. The length and width of the image are m1 and n1 pixels.

[0096] 2) Based on the coordinate system and image size parameters, randomly generate a cropping rectangle parameter set, including center point coordinates, area percentage, and aspect ratio; based on the cropping rectangle parameter set, calculate the size and boundary coordinates of the cropping rectangle.

[0097] In this embodiment, P, S, and r are a set of parameters of the cropping rectangle that are randomly generated.

[0098] The random center point position of the cropping rectangle R is P(x, y), where x≤m1, y≤n1.

[0099] The random area percentage S of the cropping rectangle, S∈[0.2,0.8], and the random aspect ratio a / b=r, r∈[0.5,2], where

[0100] The range of the cropping rectangle, the coordinates of the upper left corner are The coordinates of the lower right corner are

[0101] 3) Determine whether the cropping rectangle is completely within the original single image range. If so, store the cropping rectangle parameter set in the valid parameter set. If not, regenerate the cropping matrix parameter set.

[0102] In this embodiment, in order to ensure that the cropping rectangle does not exceed the range of the original image, the following four judgments need to be met at the same time:

[0103]

[0104] If the conditions of formulas (1), (2), (3), and (4) are not met, the cropping rectangle cannot be generated. Repeat steps 1) and 2) and randomly generate P, S, and r again until the above conditions are met.

[0105] If the conditions of formulas (1), (2), (3), and (4) are met, the cropping rectangle parameters P, S, and r are stored in the valid parameter set U.

[0106] 4) Compare the cropping matrix parameter group with the parameter groups in the valid parameter group set. If they are the same, regenerate them until a predetermined number of unique parameter groups are obtained.

[0107] In this embodiment, a new set of P, S, r is randomly generated and compared with each existing set of cropping rectangle parameters P, S, r in the set U. If they are identical, a new set of P, S, r is generated until it is different from all the rectangle parameters in the set U. Step 3) is then repeated until the number of cropping rectangle parameters in the set U reaches num.

[0108] 5) The original single images are batch-cropped according to a predetermined number of unique parameter groups to generate a predetermined number of fragmented images.

[0109] The original image is cropped according to each set of cropping rectangle parameters P, S, r in the set U to generate a cropped image, and the cropped image is added to the same type of paleontological fossil image set in the paleontological fossil image database.

[0110] Repeat the above steps to obtain a predetermined number of fragmented images of all the single images in sequence.

[0111] If there are N individual images of this type of ancient organism, then the number of fragmented images after cropping is N*num, and the number of individual and fragmented images is N*(num+1).

[0112] Finally, monomers and fragments are classified and labeled.

[0113] In this embodiment, in order to extract multi-scale life features from paleontological images and identify possible life fragments and traces in the returned samples, the images in the paleontological fossil database are randomly cropped and paleontological fossil fragment images are added to the database. Figure 4 As shown in the figure, a single paleontological image is randomly cropped five times to obtain five fragmented images. The first image is a single paleontological image, and the next five images are fragmented images after random cropping.

[0114] The images of individual paleontological fossils were cropped and added to the database, and eventually the database contained more than 150,000 images of individual paleontological fossils and fragments.

[0115] Through the above steps, the number of paleontological images has been increased, and a large number of fragment images of different parts have been generated. To solve the problem of possible fragmentation of traces of life in samples returned from Mars and other exoplanets, the number of images in the paleontological image database is expanded based on random cropping and other methods, which fragments the paleontological fossils in the library and re-standardizes the fragmented paleontological fossils. A database of both whole and local fragments of Earth paleontological fossils is established to meet the requirements for identifying traces of life in samples returned from exoplanets. This improves the number and quality of images in the Earth paleontological image database, providing a comprehensive and effective data foundation for subsequent model training and identification.

[0116] (2) Classify and label multi-category and multi-body images.

[0117] For images containing multiple species and multiple paleontological individuals, they are not cropped. The paleontological fossil individuals are classified and labeled, and the remaining rock fragments and matrix are automatically classified into one category, namely the background category.

[0118] like Figure 1 As shown, in step S2, the constructed paleontological life feature recognition model is trained based on the preprocessed images.

[0119] S2-1. Construct a model for identifying the life characteristics of ancient organisms.

[0120] In this embodiment, the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block.

[0121] The basic network module mainly includes four scale transformation blocks, each of which includes two convolutional layers.

[0122] The multi-scale feature block consists of multiple groups of attention modules and scale transformation blocks that are alternately connected in series.

[0123] The feature fusion module is a bidirectional pyramid network, which includes a bidirectional cross-scale connection structure, a depth-wise separable convolutional layer, a weighted fusion mechanism module and a depth-wise separable convolutional layer.

[0124] S2-2. Training of paleontological life characteristics identification model.

[0125] In this example, all images in the database are divided into training and test sets in a ratio of 4:1. The input images are resampled to 256*256*3, where 3 is the number of channels in the input image. The model is iteratively trained for 1000 batches.

[0126] Generate features through basic network blocks Figure 1 , then the feature Figure 1 This is input into the multi-scale feature block. The convolutional layer output is re-encoded using the attention mechanism and passed through a feature fusion module, allowing each layer to incorporate information from multiple scales. Different layers represent features at different scales of the image. Ultimately, six feature maps of different scales are output.

[0127] Taking the input image 256*256*c1 as an example, where c1=3 is the number of channels of the image, the 6th layer obtains the feature map output of 1*1*c2, where c2=768.

[0128] The constructed paleontological life feature recognition model is trained. The specific process is as follows:

[0129] S2-2-1. Extract the first feature map of the preprocessed image through the basic network module.

[0130] The preprocessed image is input into the basic network module for feature extraction to obtain the first feature map.

[0131] Input the test set images into the basic network module.

[0132] Specifically, (1) the input image I is resampled into a 3-channel image I1 of size c*c, where c is a power of 2, and n>=8.

[0133] (2) Input the channel image I1 into the basic network block and obtain the first feature map through multiple scale transformation blocks.

[0134] Specifically, the channel image I1 is sequentially input into m serial scale transformation blocks B, and the number of channels is gradually doubled. m is greater than or equal to 3 and less than or equal to n-3. After the basic network block, the output feature map Fmap1 has a size of (c / 2 m )*(c / 2 m ), the number of channels is 3*2 m Each scale transformation block B consists of two 3*3 convolutional layers with padding of 1 and a 2*2 max pooling layer with a stride of 2. Therefore, each pixel in its output image has a 6*6 receptive field of its input image. The height and width halving block expands the receptive field of each unit in its output feature map.

[0135] In this embodiment, the basic network block has a total of 4 half-length and half-width blocks, and the generated features Figure 1 The size is 16*16*48, where 48 is the number of channels.

[0136] S2-2-2. Input the first feature map into the multi-scale feature block, and perform feature enhancement through the attention module and scale transformation block.

[0137] The scale feature block consists of multiple groups of attention modules and scale transformation blocks that are alternately connected in series. The specific process is as follows:

[0138] In this embodiment, there are n groups of attention modules and scale transformation blocks that are alternately connected in series, and there are n attention modules and n-1 scale transformation blocks.

[0139] (1) The feature map is input into the first attention module for spatial feature enhancement to obtain the first enhanced feature map.

[0140] The Coordinate Attention (CA) module is introduced to capture long-range contextual information. Unlike Channel Attention, which converts the input into a single feature vector through 2D global pooling, CA decomposes channel attention into two 1D feature encoding processes that aggregate features along different directions. The advantage of this is that it can capture long-range dependencies along one spatial direction while retaining precise position information along the other spatial direction. The generated feature maps are then encoded separately to form a pair of direction-aware and position-sensitive feature maps, which can be applied complementary to the input feature map to enhance the representation of the object of interest.

[0141] Use the attention mechanism to re-encode the output of the convolutional layer to obtain the first enhanced feature map. The specific process is:

[0142] 1) Perform one-dimensional global average pooling on the first feature map along the height and width to generate height direction features and width direction features;

[0143] In this embodiment, the feature map X is used as input to the CA attention mechanism module, where: C is the number of channels, representing the different feature channels in the feature map; H is the height, representing the vertical dimension of the feature map; W is the width, representing the horizontal dimension of the feature map.

[0144] Global pooling is commonly used for channel attention. It reduces the spatial dimension of feature maps, thereby reducing parameters and computation. However, it is difficult to preserve position information. To enable the attention module to capture long-range spatial dependencies with precise position information, global pooling is decomposed into a pair of one-dimensional feature encoding operations.

[0145] Specifically, a pooling kernel of size (H, 1) and (1, W) is used to encode each channel along the horizontal and vertical coordinate directions for X. For the output of the c'th channel with a height of h', the formula is:

[0146]

[0147] in, Represents the height-direction features of c' channels.

[0148] The output of the c'th channel with a width of w' is:

[0149]

[0150] in, Represents the width-wise feature of the c'th channel.

[0151] These two feature maps capture the global features in the width and height directions respectively.

[0152] 2) After concatenating the height-direction features and the width-direction features, the dimensionality is reduced through convolution to obtain the intermediate features.

[0153] In this embodiment, z h′ and z w′ Perform a concatenation operation to combine the width and height features. Apply the convolution transformation function to the combined feature layer and perform the F1 operation, that is, use a 1*1 convolution kernel to reduce the dimension to capture the relationship between the width and height dimensions. Then, the activation function is used to further process the features and generate the feature map f. The formula is:

[0154] f=δ(F1([z h′ , z w′ ])) (7)

[0155] Among them, f represents the intermediate feature, z □′ Represents the height direction feature, z w′ Indicates width-wise features.

[0156] 3) Decompose and transpose the intermediate features to obtain width features for restoring width and height features for restoring height.

[0157] In this embodiment, f is decomposed into two separate tensors along the spatial dimension, and the features in the width and height directions are separated from the above feature layer. Then the two separated feature layers are transposed to restore the dimensions of width and height, and the feature layer f is obtained. h′ and f w′ , that is, we get the width feature of restored width and the height feature of restored height respectively.

[0158] 4) Convolution is performed on the width features of the restored width and the height features of the restored height, and the width and height attention weights are generated through the Sigmoid activation function.

[0159] In this embodiment, the feature layer f h′ and fw′ Apply 1*1 convolution transformation respectively to increase the dimension to a tensor with the same number of channels as the input. Then, apply the Sigmoid activation function to obtain the attention vector g in the width and height dimensions h′ and g w′ :

[0160] g h′ =σ(F h (f h′ )) (8)

[0161] g w′ =σ(F w (f w′ )) (9)

[0162] Among them, g h' represents the height attention weight, g w′ represents the width attention weight, f w′ The width feature representing the recovered width, f h′ Height feature representing the recovered height.

[0163] That is, they serve as height and width attention weights, respectively, to indicate the importance of different positions.

[0164] 5) Multiply the wide height and width attention weights with the multi-scale feature map to obtain the enhanced target features.

[0165] Multiply the original input feature map X by the attention weights in the width and height directions to obtain the feature map y' of the CA attention mechanism. The numerical formula of y' at height i, width j, and channel c' is:

[0166]

[0167] in, is the height attention weight g h' The vector decomposed along channel c' is, is the width attention weight g w′ The vector obtained by decomposing along channel c'.

[0168] (2) The first enhanced feature map is scaled by a scale transformation block to obtain a first scale feature map.

[0169] In this embodiment, there are n-1 scale transformation blocks. The series scale transformation blocks gradually double the number of channels. The enhanced target features are input to the scale transformation block, and the output feature map Fmap1 has a size of (c / 2 n )*(c / 2 n), the number of channels is 3*2n. Each scale transformation block B consists of two 3*3 convolutional layers with padding of 1 and a 2*2 maximum pooling layer with a stride of 2.

[0170] This module takes the first feature map Fmap1 of the basic network block as input, and uses the height and width halving block B to continuously generate new feature maps, so that the receptive field of each unit in the feature map becomes larger. The number of height and width halving blocks B in the multi-scale feature block is n-1. This means that the minimum size of the feature map generated by the input image is 1*1. After n attention modules and n-1 scale transformation blocks, the size of the final output feature map Fmapn is (c / 2 m+(n-1) )*(c / 2 m+(n-1) ), channel 3*2 m+(n-1) .

[0171] (3) The first scale feature map is input into the second attention module for spatial feature enhancement to obtain a second enhanced feature map.

[0172] (4) The second enhanced feature map is scaled by a scale transformation block to obtain a second scale feature map.

[0173] (5) Based on the above steps, iteratively input each group of attention modules and scale transformations in turn until feature maps of all scales are obtained.

[0174] Specifically, it is sequentially input into n attention modules and n-1 scale transformation blocks to obtain n enhanced feature maps, namely the first enhanced feature map, the second enhanced feature map, the third enhanced feature map, until the nth enhanced feature map is obtained.

[0175] In this embodiment, the feature Figure 1 Input into the multi-scale feature block, a total of 4 length and width halving blocks (scale transformation blocks) and attention modules, respectively generate features Figure 2-5 , the corresponding sizes are 8*8*96, 4*4*192, 2*2*384 and 1*1*768, and the last dimension is the number of channels.

[0176] S2-2-3. The enhanced multi-scale feature map is fused through the feature fusion module; based on the fused feature map, multi-position and multi-scale anchor frames are generated for the enhanced multi-scale feature map.

[0177] A weighted bidirectional pyramid structure module, BiFPN, has been added to the feature fusion network structure to learn different input features to better handle multi-scale feature fusion. Traditional feature fusion networks extract high-level pyramid features from the base network and directly predict them, but this structure does not perform feature fusion and has low accuracy. BiFPN introduces bidirectional connections, allowing information to propagate in both directions between different resolution levels. This helps to better fuse low-level and high-level features and promotes the contextual propagation of features, thereby improving the accuracy of object detection and segmentation. At the same time, an adaptive feature adjustment mechanism can learn weights to adjust features at different levels to better match the needs of different tasks.

[0178] Traditional feature fusion often simply superimposes or adds feature maps without distinguishing between them. However, different input feature maps have different resolutions and contribute differently to the fused input feature map. Therefore, simply superimposing or adding them is not optimal. Therefore, BiFPN uses a simple yet efficient weighted fusion mechanism.

[0179] The feature fusion module is a bidirectional pyramid network, which includes a bidirectional cross-scale connection structure, a depth-wise separable convolutional layer, a weighted fusion mechanism module and a depth-wise separable convolutional layer.

[0180] (1) The enhanced multi-scale feature maps are fused through the feature fusion module. The specific process is as follows:

[0181] 1) In a bidirectional pyramid network, bidirectional feature propagation from top to bottom and from bottom to top is performed through a bidirectional cross-scale connection structure;

[0182] BiFPN introduces bidirectional connections, allowing information to propagate bidirectionally between different resolution levels. This helps to better fuse low-level and high-level features and promotes the contextual propagation of features, thereby improving the accuracy of object detection and segmentation.

[0183] 2) The multi-scale feature maps are weightedly fused through the weighted fusion mechanism module to obtain different scale feature map outputs;

[0184] Simply superimposing or adding is not the best operation. In this embodiment, BiFPN uses a simple and efficient weighted fusion mechanism, and the formula is:

[0185]

[0186] Among them, Output is the output, which means the weighted sum of all inputs, w i and w jThey represent the parameter values ​​learned from the i-th input and the j-th input during the training process, respectively. ∈ is a small value, ∈=0.0001, which is used to stabilize the output range and prevent large fluctuations in various parameters during the learning process. Different values ​​can be taken according to actual conditions.

[0187] In each w i Then add the linear rectification activation function to ensure that w i ≥0, that is, w i =ReLU(w i ), and use small values ​​∈ to avoid numerical instability. Thus, based on bidirectional cross-scale connections and normalized fusion, the final weighted bidirectional feature pyramid network is obtained.

[0188] 3) Each scale feature output is processed through a depth-wise separable convolutional layer to obtain a processed multi-scale feature map, which is then downsampled or upsampled through a resize operation to obtain an aligned multi-scale feature map; the aligned scale feature maps are fused using a normalized fusion method to obtain a fused feature map.

[0189] by Figure 2 Taking the output of the P6 layer as an example, the calculation expression is:

[0190]

[0191] in, It is the intermediate feature of the sixth layer in the top-down path; It is the output feature of the sixth layer in the bottom-up path; Conv is a depth-separable convolution, and normalization and activation functions are added after each convolution; Resize operation is a downsampling or upsampling operation;

[0192] w1, w2, w1, w2, w3 are respectively Figure 2 The parameters learned during the learning process in the five directions indicated by the arrows are used to distinguish the importance of different features in the feature fusion process; and are the input features of the sixth and seventh layers respectively; is the output feature of the fifth layer in the bottom-up path.

[0193] (2) Based on the fused feature map, multi-position and multi-scale anchor frames are generated for the enhanced multi-scale feature map.

[0194] The anchor box generation block AFB is used to generate multi-position and multi-scale anchor boxes AF for the enhanced multi-scale feature map.

[0195] During training, multi-scale anchor boxes are generated during the model's forward propagation. The generated anchor boxes are then annotated with categories and offsets based on the label information. Finally, a loss function is calculated based on the predicted and annotated values ​​of the categories and offsets.

[0196] For each feature map, anchor boxes are generated. The scale ratio is chosen by evenly dividing the range from 0.2 to 1.05 into five parts, corresponding to scale ratios of 0.2, 0.37, 0.54, 0.71, 0.88, and 1.05, and aspect ratios of 1, 2, and 0.5. All combinations of these aspect ratios and scale ratios are used, centered around each pixel in the feature map Fmap.

[0197] In this embodiment, multi-position and multi-scale anchor frames (i.e., rectangular frames) are generated for each feature map. Assume that the height of one feature map Fmap2 is h and the width is w. Anchor frames of different shapes are generated with each pixel of the feature map as the center. The scaling ratio of the anchor frame is e∈(0,1] and the aspect ratio is f>0. Then the width and height of the anchor frame are and

[0198] In order to generate multiple anchor boxes of different shapes at one pixel position, the scaling ratios are set to e1, e2, ..., e i ,i=1,2,…,m2,the aspect ratio is f1,f2,…,f j , j = 1, 2, ..., n2. Using all combinations of these aspect ratios and scaling ratios with each pixel of the feature map Fmap2 as the center, the input feature map will have a total of w*h*m2*n2 anchor boxes. The size of the anchor box generation block AFB is the same as the size of the corresponding feature map. For example, the size of the fifth anchor box generation block AFB5 is 1*1.

[0199] Specifically, each feature map that fuses multi-scale features is output by the feature fusion module. Assume that the height of the nth feature map Fmapn' is h n , with a width of w n . Fmapn' is input to the nth anchor frame to generate block AFBn. The scaling ratio value is evenly divided into 5 parts between 0.2-1.05, and the corresponding scaling ratios are 0.2, 0.37, 0.54, 0.71, 0.88 and 1.05, and the aspect ratios are 1, 2 and 0.5. For each pixel of Fmapn' as the center, all combinations of these aspect ratios and scaling ratios are used, that is, the block AFBn is generated by the anchor frame to obtain the nth anchor frame set AFUn, the size of the set is w n *h n *5*3, each element in the anchor frame set AFUn corresponds to a pixel in the feature map Fmapn', and the size of the anchor frame generated with the pixel as the center.

[0200] S2-2-4. Perform target classification and position prediction on the image within the anchor box through the category prediction and position prediction blocks.

[0201] The image within the anchor box is used for target classification and location prediction.

[0202] In the training set, each anchor box is considered a training example. Therefore, in order to train the object detection model, each anchor box is labeled with its category and offset. The former is the category of the object associated with the anchor box, and the latter is the offset of the ground truth bounding box relative to the anchor box.

[0203] During the training process, to determine which anchor box to match the true bounding box of the category in the training image, there are two matching principles.

[0204] The first principle is: for each real bounding box in the image, find the anchor box with the largest intersection-over-union (IOU) and match it.

[0205] The second principle is: for the remaining unmatched anchor boxes, if the IOU with a ground-truth bounding box is greater than a threshold T (usually 0.5), then the anchor box is also matched with the ground-truth bounding box. The IOU is the ratio of the area of ​​the intersection and union of the anchor box and the ground-truth bounding box.

[0206] Implement the first principle first, then the second.

[0207] It is divided into two blocks: the category prediction block and the position prediction block, which are used for target classification and position regression of the anchor box content, respectively. Each block contains two 3*3 convolutional layers with padding of 1, so that the height and width of the input and output remain unchanged, but the number of output channels of the convolutional layers of the two blocks is different. Assuming that there are q types of fossil categories in the paleontological database, then together with the non-life feature categories in the micrograph (i.e., the background, that is, rock debris or matrix other than life features, is classified as one category), there are q+1 categories in total. For example, the input is the feature map Fmap3, and the output is the feature map Fmap_C for category prediction and the feature map Fmap_L for position prediction. Because the coordinate channel of the output category feature map Fmap_C contains the category predictions of all anchor boxes corresponding to the coordinates of the output feature map, the number of output channels is a*(q+1), where a is the total number of anchor boxes on the feature map, a=w*h*n*m.

[0208] Similarly, the position prediction block predicts four offsets for each anchor box: the horizontal and vertical offsets between the center point of the anchor box and the center point of the true category bounding box, as well as the width and height deviations between the anchor box and the true category bounding box. Therefore, for the input feature map Fmap3, the output channel of the position prediction block is a*4.

[0209] S2-4. Construct a loss function optimization model.

[0210] (1) Loss of anchor box category

[0211] There are two types of loss in detecting life features in paleontological fossils. The first is the anchor box category loss, which is the loss between the life feature type (paleontological fragment category) predicted in the anchor box generated in the feature graph and the true category. It is calculated using the cross-entropy loss function FLoss. FLoss is based on the cross-entropy loss of binary classification and is a dynamically scaled cross-entropy loss. It can dynamically reduce the weight of easily distinguishable samples during training, thereby quickly focusing on those difficult-to-distinguish samples. The formula for the binary cross-entropy loss function is:

[0212]

[0213] Among them, y represents the true label of the sample, Represents the label predicted by the model. When y = 1, it means the sample belongs to the positive class; when y = 0, it means the sample belongs to the negative class.

[0214] In practical problems, since the loss of a single easy-to-separate sample is smaller than the loss of a single hard-to-separate sample, if the number of easy-to-separate samples is much greater than that of hard-to-separate samples, the loss of all samples may be dominated by the loss of a large number of easy-to-separate samples, resulting in the inability to fully learn the hard-to-separate samples. FLoss considers the problem of imbalance between easy-to-separate samples and introduces a modulation factor (1-p t ) γ , where 1-p t ∈[0,1],γ≥0, and weight factor α t , α∈[0,1], when it is a positive sample, the weight factor is α, when it is a negative sample, the weight factor is 1-α, the class label of the current sample is t, p t represents the probability value predicted by the classifier as class t, and the loss function is defined as:

[0215] FLoss(p t )=-α t (1-p t ) γ log(p t ) (15)

[0216] The modulation factor is used to reduce the loss contribution of easy-to-separate samples, whether foreground or background, p t The larger it is, the easier it is to distinguish the sample and the smaller the modulation factor is. t Used to adjust the ratio between positive and negative sample losses, the foreground category uses α t When the corresponding background category uses 1-α t .

[0217] (2) Loss of anchor frame position offset

[0218] The second is the loss of the offset of the anchor frame of the paleontological fossil fragment, that is, the loss of the offset between the predicted anchor frame position of the paleontological fossil fragment and the true anchor frame. The L1 norm loss (L1Loss) can be used to calculate the absolute difference between the predicted value and the true value. The L1 norm loss function is to convert the target value (Y i ) and the estimated value (f(x i )) is minimized, the formula is:

[0219]

[0220] Finally, the final loss function of the model is F'=FLoss+L1Loss, which is the sum of the losses of the anchor box category and the anchor box position offset.

[0221] The model is optimized through the final loss function.

[0222] The results of data verification in the test set are as follows Figure 5 As shown in the figure, the colored boxes represent the locations of life features, the names represent the most likely paleontological species, and the numbers represent the highest probability of being identified as that paleontological species. It can be seen that the fragments of paleontological remains can be correctly identified, with a recognition rate of over 90% for images in the database.

[0223] like Figure 1 As shown, in step S3, the trained paleontological life feature recognition model is used to perform life feature recognition on the paleontological fragment image to obtain a recognition result.

[0224] The model for extracting paleontological life signatures is being applied to identifying traces of life in samples returned from exoplanets. However, the Tianwen-3 Mars sampling mission is scheduled to launch in 2028, with samples expected to return around 2030. Preliminary research to identify traces of life must begin before the samples return, and it is impossible to wait until the samples return to begin database construction and life signature extraction. Therefore, at this stage, we will use microscopic images of paleontological remains in Earth's sedimentary rocks as an example to develop model predictions.

[0225] In this embodiment, the image of the paleontological fragment to be identified is input into the paleontological life feature recognition model. The specific process is as follows:

[0226] (1) Read the microscopic image I to be predicted and resample it into a c*c image I'; extract the first feature map of the image I' through the basic network.

[0227] Then, an enhanced multi-scale feature map is generated through the multi-scale feature block; then, the fused feature map is obtained through the feature fusion module, and based on the fused feature map, an anchor box is generated for each pixel of each feature map.

[0228] (2) For each anchor box, use the category prediction block to predict and generate the corresponding category confidence. Determine its category (the one with the highest confidence) and confidence value based on the category confidence, and filter out anchor boxes belonging to the background. Filter out anchor boxes with lower confidence based on the confidence threshold (e.g., 0.5), sort them in descending order based on confidence, and then retain only the top anchor boxes (e.g., top = 400). Use the position prediction block on the retained anchor boxes to obtain the predicted bounding box of the category with the highest confidence.

[0229] Non-maximum suppression is used to remove similar predicted bounding boxes, and all anchor boxes with a confidence level of not less than 0.9 are selected as the predicted bounding boxes of the corresponding category. In the end, only the predicted bounding boxes that meet specific conditions are output.

[0230] like Figure 6 As shown in the figure, images of paleontological fragments collected outside the Earth paleontological database are fed into a trained paleontological life signature extraction and recognition model for prediction. This method can clearly identify individual paleontological entities and fragments in the image. Different colored boxes circle the traces of life on the slides, and the most likely fossil species and probability values ​​are given. Therefore, this method can accurately identify traces of paleontological life in microscopic images and could be used in the future to search for traces of life in samples returned from exoplanets, providing a powerful auxiliary tool for deep space exploration.

[0231] The model in this example is suitable for searching for traces of life in samples returned from future missions to asteroids and exoplanets (Mars, Jupiter's moons, Venus, etc.). It is trained and uses images of Earth's paleontological fossils to extract life signatures, thereby identifying and extracting paleontological fragments or traces of life in microscopic images of samples returned from exoplanets.

[0232] Example 2

[0233] The purpose of this embodiment is to provide a deep learning-based system for identifying life traces on extraterrestrial planets, including:

[0234] A database construction and data preprocessing module is used to construct a paleontological fossil image database and preprocess the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-class multi-body images;

[0235] A model training module is used to train a constructed paleontological life feature recognition model based on preprocessed images; wherein, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is as follows: extracting a feature map of the preprocessed image through the basic network module; inputting the feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map based on the fused features; and performing target classification and position prediction on the image within the anchor frame through the category prediction and position prediction blocks;

[0236] The model recognition module is used to use the trained paleontological life feature recognition model to perform life feature recognition on the paleontological fragment images to obtain recognition results.

[0237] Based on providing a deep learning-based identification system for extraterrestrial life traces, the method steps in Example 1 are implemented.

[0238] Example 3

[0239] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0240] Example 4

[0241] The purpose of this embodiment is to provide a computer-readable storage medium.

[0242] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0243] The steps involved in the apparatus of the above embodiment correspond to those of the method embodiment 1. For detailed implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0244] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0245] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for identifying life traces on extraterrestrial planets based on deep learning, characterized in that: include: Constructing a paleontological fossil image database and preprocessing the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-category multi-body images; Based on the pre-processed images, the constructed paleontological life feature recognition model is trained; Using the trained paleontological life feature recognition model, the paleontological fragment images to be identified are used to identify life features and obtain recognition results; Among them, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is: extracting the first feature map of the preprocessed image through the basic network module; inputting the first feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; based on the fused feature map, generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map; and performing target classification and position prediction on the image in the anchor frame through the category prediction and position prediction blocks.

2. The method for identifying life traces on extraterrestrial planets based on deep learning according to claim 1, characterized in that: Construct a paleontological fossil image database. Specifically, according to the collection rules, collect images of Earth paleontological fossils from different eras related to extraterrestrial planets, and annotate and accurately classify them to obtain a paleontological fossil image database including images of protozoa, lower animals and plant pollen fossils.

3. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 2, characterized in that: The collection rules include: Select molded fossils and chemical fossils as fossil types, and collect fossils of ancient Earth organisms observed under a microscope at millimeter to micrometer levels; Collecting spores and pollen fossils of protists, lower animals and plants on Earth; The collected fossils of reef-building organisms include individuals and complexes.

4. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 1, characterized in that: Preprocessing of paleontological fossil images, the specific process is as follows: Fragmenting the single image to obtain a fragmented image; Classifying and labeling the single image and the fragmented image; Classify and label multi-class and multi-body images.

5. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 4, characterized in that: The fragmentation process of the single image refers to randomly cropping the single image of the paleontological fossil by a rectangular shape. The specific process is as follows: Setting a coordinate system with the lower left corner of the single image as the origin and determining the image size parameters; Based on the coordinate system and image size parameters, a cropping rectangle parameter set is randomly generated, including the center point coordinates, area percentage, and aspect ratio; Calculating the size and boundary coordinates of the cropping rectangle according to the cropping rectangle parameter group; Determine whether the cropping rectangle is completely within the original single image range. If so, store the cropping rectangle parameter group in the valid parameter group set. If not, regenerate the cropping matrix parameter group. Comparing the clipping matrix parameter group with the parameter groups in the valid parameter group set, and regenerating them if they are the same, until a predetermined number of unique parameter groups are obtained; Batch cropping the original single image according to a predetermined number of unique parameter groups to generate a predetermined number of fragmented images; Repeat the above steps to obtain a predetermined number of fragmented images of all the single images in sequence.

6. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 1, characterized in that: The first feature map is input into a multi-scale feature block, and feature enhancement is performed through an attention module and a scale transformation block. The multi-scale feature block includes multiple groups of attention modules and scale transformation blocks alternately connected in series. The specific process is as follows: Inputting the first feature map into a first attention module for spatial feature enhancement to obtain a first enhanced feature map; Performing a scale transformation on the first enhanced feature map by a scale transformation block to obtain a first scale feature map; Inputting the first scale feature map into the second attention module for spatial feature enhancement to obtain a second enhanced feature map; Performing a scale transformation on the second enhanced feature map by a scale transformation block to obtain a second scale feature map; Based on the above steps, the input is iterated to each group of attention modules and scale transformations in turn until feature maps of all scales are obtained.

7. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 1, characterized in that: The first feature map is input into the first attention module for spatial feature enhancement. The specific process is as follows: Performing one-dimensional global average pooling on the first feature map along the height and width respectively to generate height-direction features and width-direction features; After concatenating the height-direction features and the width-direction features, the dimensionality is reduced through convolution to obtain the intermediate features. Decompose and transpose the intermediate features to obtain width features for restoring width and height features for restoring height respectively; The width feature of the restored width and the height feature of the restored height are convolved and dimensioned respectively, and the width and height attention weights are generated through the Sigmoid activation function; The width and height attention weights are multiplied with the multi-scale feature map to obtain enhanced target features.

8. The method for identifying life traces on extraterrestrial planets based on deep learning as claimed in claim 1, characterized in that: The multi-scale feature maps are fused through a feature fusion module. Among them, the feature fusion module adopts a bidirectional pyramid network, including a bidirectional cross-scale connection structure, a depth-separable convolutional layer, a weighted fusion mechanism module and a depth-separable convolutional layer. The specific process is as follows: In the bidirectional pyramid network, bidirectional feature propagation from top to bottom and from bottom to top is performed through a bidirectional cross-scale connection structure; The multi-scale feature maps are weightedly fused through the weighted fusion mechanism module to obtain different scale feature map outputs; Each scale feature output is processed through a depth-wise separable convolutional layer to obtain a processed multi-scale feature map; The processed multi-scale feature map is downsampled or upsampled through the Resize operation to obtain the aligned multi-scale feature map; The normalized fusion method is used to fuse the aligned scale feature maps to obtain the fused feature maps.

9. A deep learning-based system for identifying life traces on extraterrestrial planets, characterized by: include: A database construction and data preprocessing module is used to construct a paleontological fossil image database and preprocess the paleontological fossil images, wherein the image formats in the paleontological fossil image database are: single-body images and multi-class multi-body images; A model training module is used to train a constructed paleontological life feature recognition model based on preprocessed images; wherein, the constructed paleontological life feature recognition model is trained, and the paleontological life feature recognition model includes a basic network module, a multi-scale feature block, a feature fusion module, and a category prediction and position prediction block. The specific process is as follows: extracting the first feature map of the preprocessed image through the basic network module; inputting the first feature map into the multi-scale feature block, and performing feature enhancement through the attention module and the scale transformation block; fusing the enhanced multi-scale feature map through the feature fusion module; based on the fused feature map, generating multi-position and multi-scale anchor frames for the enhanced multi-scale feature map; and performing target classification and position prediction on the image within the anchor frame through the category prediction and position prediction blocks; The model recognition module is used to use the trained paleontological life feature recognition model to perform life feature recognition on the paleontological fragment images to obtain recognition results.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images