Processing method and system based on computer vision large model

Abstract

The invention relates to the technical field of computer vision, in particular to a processing method and system based on a computer vision large model, and the method comprises the following steps: obtaining the alignment position of a continuous response region in a feature map, screening a consistent change region, analyzing semantic nodes, connecting a mapping path to an original map, extracting coordinate analysis direction grouping nodes, and tracking semantic change to mark a jump position, extracting feature analysis to activate a difference to delimit a mutation fragment, and obtaining a semantic response structure chart. According to the method, continuous regions with consistent directions are screened through position alignment of multilayer channel response regions, spatial focusing of semantic regions is carried out, path clues with continuous semantic relationships are screened through extraction of connection directions and sequence information of nodes in an output sequence, and path segments with a propelling trend are classified through analysis of node coordinate arrangement. And by tracking node semantic value changes, marking jump node positions, and dividing mutation fragments through response differences, the layering and distinguishing capability of the semantic structure is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a processing method and system based on a computer vision large model. Background Art

[0002] The field of computer vision encompasses technical solutions for analyzing, identifying, understanding, and reasoning based on image data. The core of this technical field is the use of image acquisition devices to acquire image information and, through a series of processing steps, automatically identify, classify, detect, and understand the image content. Computer vision is widely used in tasks such as image recognition, object detection, behavior recognition, and scene analysis. Its key lies in converting the information contained in an image or video into a structured representation that can be understood and processed by a computer. In specific applications, computer vision is often combined with deep learning models to construct an algorithmic system with semantic understanding capabilities, thereby enabling multi-level abstract analysis of image content.

[0003] Among them, the processing method based on a large computer vision model refers to a processing method that uses a pre-trained large-scale model with visual language fusion capabilities to perform multimodal understanding and representation reconstruction on image input information. The technical matters covered by this patent subject include using a multimodal pre-trained model to semantically encode image content, extracting and aggregating the image semantic layer by introducing a text-guided mechanism, and processing the input image in combination with a cross-modal feature alignment mechanism, thereby outputting structured visual features in a fixed format. Its processing method generally includes using a large-scale visual language model as the basic architecture, performing steps such as feature embedding, cross-modal fusion, semantic-guided alignment, and feature remapping on the image input, thereby achieving unified encoding and processing of image content.

[0004] When processing image semantic structures, existing technologies often integrate different semantic regions into the same expression space, ignoring the spatial expression requirements for feature continuity between regions. Path information is often implicitly generated by the model, lacking assessment of directional stability and sequential integrity. This leads to a mixed expression of semantic structures in the image and a flattened distribution of structural paths. Semantic mutation points are often not independently extracted in the processing logic, and response segment division relies on the overall channel response intensity, making it difficult to capture local differences. Due to frequent response overlap, some boundary regions lack a targeted structural differentiation mechanism, forming fuzzy boundary bands, which affect the clarity and accuracy of subsequent structured expression. For example, in multi-target images, if there is directional overlap between targets or densely connected response regions, current solutions cannot effectively split the path structure, resulting in a loss of the organizational hierarchy between semantics in the atlas and prone to problems such as label drift and response interference. These problems limit the sophistication and directional constraints of semantic information at the structural expression level, reducing the stability of the overall semantic parsing results. Summary of the Invention

[0005] In order to solve the technical problems existing in the prior art, an embodiment of the present invention provides a processing method based on a large computer vision model, comprising the following steps: In order to achieve the above object, the present invention adopts the following technical solution: a processing method based on a large computer vision model, comprising the following steps: S1: Obtain the feature map of the input image frame through the multi-layer feature extraction structure, extract the continuous response area of ​​the same position in the channel, align the cross-layer area positions, filter the areas with consistent change directions, and obtain the set of focused feature areas; S2: Based on the set of focused feature regions, extract the semantic information nodes of the regions in the output sequence, analyze the node connection direction and semantic order, screen the path clues with consistent directions, and map the path positions to the original image regions to obtain the regional semantic dependency path sequence; S3: Based on the regional semantic dependency path sequence, extract the spatial coordinates of the nodes in the path, analyze the consistency of the arrangement directions of adjacent coordinates, group the nodes with continuous advancement directions into one group, and obtain a semantic guidance path set; S4: Based on the semantic guidance path set, tracking the change trend between the semantic values ​​of the nodes in the path, marking the jump nodes with position offsets, recording the corresponding coordinates, and obtaining a semantic jump node position set; S5: Based on the semantic jump node position set, extract the local features of the nodes in the channel response structure, analyze the spatial arrangement and activation area differences between adjacent nodes, delineate the mutation segments, and obtain the semantic response area distribution structure diagram.

[0006] As a further solution of the present invention, the focused feature area set includes continuous response areas, inter-layer alignment information, and cross-layer consistency labels; the regional semantic dependency path sequence includes semantic information points, connection order, path direction, and image position annotations; the semantic guided path set includes spatial coordinate order, direction consistency sequence, and continuity node category; the semantic jump node position set includes jump node image coordinates, index numbers, and offset trend points; the semantic response area distribution structure diagram includes feature response mutation positions, channel activity distribution differences, and independent response fragments.

[0007] As a further solution of the present invention, the specific steps of S1 are: S101: Obtain multiple sets of feature maps of the input image frame through a multi-layer feature extraction structure, call the response amplitude value of the same position in the channel, compare the response amplitude change of the coordinate point according to the channel number, and select the coordinate point area with the same response amplitude change direction on the continuous channel to obtain the response continuous point group; S102: Based on the response continuation point group, call the position index value in each layer feature map, compare the position correspondence relationship of the same point between the differentiated layers, filter out the points with discontinuous cross-layer coordinate correlation relationship, and obtain the layer-aligned position group; S103: Based on the response amplitude of each position in the hierarchical alignment position group in the channel direction, and comparing the channel response direction change, screen the point areas whose response change trends remain consistent between differentiated channels, identify the continuous point distribution segments with change characteristics, and obtain a set of focused feature areas.

[0008] As a further solution of the present invention, the specific steps of S2 are: S201: Based on the set of focused feature regions, extract the semantic coding content corresponding to the regions in the output sequence, call the sequence index and spatial coordinates of each semantic node, compare the connection order between nodes according to the order of index numbers, and select node pairs with consecutive numbers to obtain a semantic connection structure set; S202: Calculating the direction angle change between adjacent path segments based on the direction vector data of each group of connection segments in the semantic connection structure set, filtering out path segments with sudden direction changes, and retaining path combinations with continuous direction change amplitudes to obtain a set of direction-continuous paths; S203: Based on the spatial coordinates of the nodes in the direction-continuous path set in the output sequence, retrieve the original image position index corresponding to the node, arrange each group of image coordinates according to the connection order, and obtain the regional semantic dependency path sequence.

[0009] As a further solution of the present invention, the calculation formula of the change value of the direction angle between adjacent path segments is specifically: ; in, Represents a path segment With path segments The change in the direction angle between Represents a path segment The direction angle, Represents a path segment The direction angle, Represents a path segment The direction vector modulus is Represents a path segment and path segments The modulus of the direction vector difference, represents the exponential decay factor.

[0010] As a further solution of the present invention, the specific steps of S3 are: S301: Based on the region semantic dependency path sequence, extract the two-dimensional coordinate information of the nodes in the path within the image frame, number the nodes according to the arrangement order in the path, record the coordinate position data of each pair of adjacent nodes, and arrange the recorded results in sequence to obtain a path node coordinate sequence set; S302: Based on the path node coordinate sequence set, calculate the horizontal displacement value of each pair of adjacent nodes, identify the spatial offset direction between adjacent nodes, analyze the frequency of direction changes in continuous node segments, mark the coordinate segments with continuous direction trends, and obtain node distribution segments with consistent directions; S303: Based on the node distribution segments with consistent directions, filter the node groups whose adjacent node sequence numbers are continuous, divide the node sequence into continuous number combinations, and list the node index groups according to the original arrangement order of the path to obtain a semantically guided path set.

[0011] As a further solution of the present invention, the calculation formula of the horizontal displacement value of each pair of adjacent nodes is specifically: ; in, Represents the path The horizontal displacement eigenvalue between a node and its adjacent nodes, Representative The horizontal coordinate of each path node, Representatives and The horizontal coordinates of the nodes adjacent to the node, Represents the distance between nodes, Representative The vertical coordinate of the path node, Representatives and The vertical coordinates of the nodes adjacent to the node, Representative The vertical coordinates of the path nodes, Representatives and The vertical coordinates of the nodes adjacent to the node, represents the terrain factor, represents the horizontal coordinate difference, represents the vertical coordinate difference, represents the vertical coordinate difference, represents the error term.

[0012] The specific steps of further solution S4 of the present invention are: S401: Based on the semantic guidance path set, analyzing the sequence of semantic feature values ​​of nodes in the path, continuously changing the semantic values ​​of adjacent nodes, identifying connected segments with different change amplitudes in the sequence, and obtaining semantic offset paragraph information; S402: Based on the semantic offset paragraph information, perform sequence comparison on the displacement changes of the path nodes along the change direction, divide the jump boundary according to the distribution density of the semantic difference between the nodes, and obtain the node jump change mark value; S403: Based on the node jump change mark value, the node positions of the offset mutation in the path structure are subjected to coordinate marking and serial numbering in the image area, and the corresponding image coordinates and numbers are calibrated in sequence according to the sequential arrangement of the nodes in the path to obtain a semantic jump node position set.

[0013] As a further solution of the present invention, the specific steps of S5 are: S501: Based on the semantic transition node position set, locate the channel region to which each node belongs in the multi-channel response structure, extract the node response amplitude item in the channel direction, and arrange them in sequence according to the channel number order and the node spatial position order to obtain a channel response mapping sequence; S502: Based on the channel response mapping sequence, compare the arrangement order of continuous nodes with the boundary information of the channel active area, identify the position group where the distribution continuity is interrupted in the spatial connection structure within the channel, and obtain characteristic distribution fracture information; S503: calling the feature distribution fracture information, locating the corresponding position of the jump node segment in the image area, extending the edge linear structure direction according to the connection boundary change direction, and obtaining the semantic response area distribution structure diagram.

[0014] A processing system based on a large computer vision model, comprising: The feature focus extraction module obtains the feature map of the input image frame through the multi-layer feature extraction structure. Based on the difference in response amplitude in each channel direction, it identifies the areas with continuous response at the same position, aligns the positions of the areas at each level, and marks the areas with consistent change characteristics across layers to obtain a set of focused feature areas. The semantic path construction module extracts the semantic information content of the corresponding area in the output sequence based on the set of focused feature areas, calls the connection direction data and semantic sequence data, analyzes the path direction consistency between semantic nodes, selects path information with consistent connection direction and continuous semantic relationship, and marks the corresponding position on the original image to obtain the regional semantic dependency path sequence; The path sequence classification module extracts the spatial coordinate order of the nodes in the path in the image based on the regional semantic dependency path sequence, compares the consistency of the arrangement direction order of adjacent coordinate points, and classifies the node sequences with prominent continuity according to the position advancement relationship to obtain a semantically guided path set; The transition node positioning module tracks the change amplitude of the semantic feature value of each node in the path between consecutive positions based on the semantic guidance path set, tracks the change position according to the offset trend, marks the image coordinates and index numbers of the points where the transition occurs, obtains the semantic transition node position set, and transmits it to the response area identification module; The response region identification module extracts the characteristic content of each node in the channel response structure based on the semantic jump node position set, compares the position arrangement between adjacent nodes with the distribution results of active areas in the channel, identifies the position segments where the characteristic response has a sudden change, and delineates them as independent response segments to obtain a semantic response region distribution structure diagram; Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, the position alignment of multi-layer channel response areas is performed to screen continuous areas with consistent directions and perform spatial focusing of semantic areas. The connection direction and order information of nodes in the output sequence are extracted to screen path clues with continuous semantic relationships. The node coordinate arrangement is analyzed to classify path segments with advancement trends. The changes in node semantic values ​​are tracked to mark the positions of jump nodes. The mutation fragments are divided according to the response differences to enhance the hierarchical differentiation ability of semantic structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 Schematic diagram of the steps of the present invention; Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0018] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0019] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same.

[0020] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0021] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0022] See also Figure 1 , an embodiment of the present invention provides a processing method based on a computer vision large model, comprising the following steps: S1: Obtain the feature map of the input image frame through the v multi-layer feature extraction structure, identify the areas with continuous responses at the same position based on the difference in response amplitude in each channel direction, align the areas between each layer, mark the areas with consistent change characteristics across layers, and obtain the set of focused feature areas; S2: Based on the set of focused feature regions, the semantic information content of the corresponding regions in the output sequence is extracted, the connection order between the semantic information is sorted out, and the path information with consistent connection direction and continuous semantic relationship is selected. The corresponding positions are marked on the original image to obtain the regional semantic dependency path sequence; S3: Based on the regional semantic dependency path sequence, the spatial coordinate order of the nodes in the path in the image is extracted, the arrangement direction order consistency of adjacent coordinate points is compared, and the node sequences with prominent continuity are classified through the position advancement relationship to obtain the semantic guidance path set; S4: Based on the semantic guidance path set, compare the change amplitude between the consecutive positions of the semantic feature value of each node in the path, track the change position according to the offset trend, mark the image coordinates and index number of the jump point, and obtain the semantic jump node position set; S5: Based on the semantic jump node position set, the characteristic content of each node in the channel response structure is extracted, and the position arrangement between adjacent nodes is compared with the distribution results of the active area in the channel. The position segments where the characteristic response has a sudden change are identified and delineated as independent response segments to obtain the semantic response area distribution structure diagram.

[0023] The focused feature area set includes continuous response areas, inter-layer alignment information, and cross-layer consistency labels. The regional semantic dependency path sequence includes semantic information points, connection order, path direction, and image position annotations. The semantic guidance path set includes spatial coordinate order, direction consistency sequence, and continuity node category. The semantic jump node position set includes jump node image coordinates, index numbers, and offset trend points. The semantic response area distribution structure diagram includes feature response mutation positions, channel activity distribution differences, and independent response fragments.

[0024] The specific steps of S1 are: S101: Obtain multiple sets of feature maps of the input image frame through a multi-layer feature extraction structure, call the response amplitude value of the same position in the channel, compare the response amplitude change of the coordinate point according to the channel number, and select the coordinate point area with the same response amplitude change direction on the continuous channel to obtain the response continuous point group; The input image frame is a single-frame static image, with the image width direction set to W, the height direction set to H, and the number of channels set to C. The input image frame first passes through a multi-layer feature extraction structure in sequence to extract a two-dimensional feature response map for each channel in each layer. The pixel position in each response map maintains a one-to-one correspondence with the original coordinates of the image. For each channel direction, the channel number is set to ci, and the response amplitude value vi (x, y) is read at the same spatial coordinate position (x, y). For each point (x, y), the response value sequence between consecutive channel numbers ci to ci+n is read in sequence and compared. The amplitude difference sign between adjacent channels is used to determine whether the response change direction remains consistent, where the change direction is set to forward or reverse. For all positions (x, y) with consistent change directions, the number of occurrences is cumulatively counted. For each point, if the number of channels with consistent continuous directions exceeds the set number threshold q, then the point is regarded as a response continuation point, and its spatial coordinates are marked and included as candidate points. All candidate points are clustered. If the number of points in a continuous area exceeds the set density threshold p, then the continuous area is identified as a continuation point area. Based on this, the edge of the point set is recorded to obtain a response continuation point group.

[0025] S102: Based on the response continuation point group, call the position index value in each layer feature map, compare the position correspondence of the same point between the differentiated layers, filter out the points with discontinuous coordinate correlation between the layers, and obtain the layer-aligned position group; Each point has the corresponding (x, y) coordinate information in the original image and its index position in different feature layers. Based on the point group, the index information of the same coordinate point in each layer of the feature map is called in turn, and each layer is numbered Li. The feature map size is set to Fi×Fi. The corresponding position of point p in each layer is pi (Li). Find pj (Lj) in the j-th layer map and record the continuous position index relationship of the point in the layers L1 to Ln. If there is a lack of response to the point position in a certain layer or the position change exceeds the offset limit range r, the point position is considered discontinuous and the point is removed. Before removal, it is necessary to compare the offsets of the positions of the two adjacent layers. If the absolute position difference exceeds the set continuous offset value setting interval [-d, +d], it is considered discontinuous. For example: a point p in L 1 is (32, 28), and L2 is (36, 29). The horizontal offset is 4 and the vertical offset is 1. If d is set to 3, the point is eliminated. The judgment condition is based on whether the absolute difference range exceeds the set value. All retained points have a continuous mapping relationship in each layer index. The index numbers of these points in different layers are combined to form a one-to-one matching table. Each row in the matching table represents the spatial coordinates and index set corresponding to each valid point between each layer. The table is further used to screen the point set that meets the continuous coordinate change trend. Skip-layer matching behavior is not allowed in the elimination operation, that is, there is no situation where a point exists in both L1 and L3 but is missing in L2. Ensure that the comparison operation is performed layer by layer in each consecutive layer. Finally, the retained point set is uniformly numbered and sorted to obtain the hierarchical alignment position group.

[0026] S103: Based on the response amplitude of each position in the hierarchical alignment position group in the channel direction, and comparing the channel response direction changes, screen the point areas where the response change trend remains consistent between the differentiated channels, identify the continuous point distribution segments with change characteristics, and obtain a set of focused feature areas; First, call the response amplitude value of each point in all channel directions, set the channel number as C1 to Cn, count the response amplitude of each point in the corresponding coordinates in different channels one by one, calculate the response change direction of the same position in adjacent channels, and use the sign change to determine the change direction. If the response of point p in C1 is 15, 18 in C2, and 22 in C3, the direction change is positive. If C2 is 12, there is a negative change. A channel direction change sequence is formed for each point, and the points with consistent direction changes in the channel sequence are further selected. The same judgment rule is applied to all points. If the change direction in the channel sequence of a certain point is continuous and consistent, they are classified into the same group. The area is formed into a preliminary point partition set, and then the set is traversed according to the proximity of spatial positions. Adjacent points are merged according to the horizontal and vertical continuous distribution of (x, y) coordinates. The judgment standard is whether the horizontal or vertical position difference is 1. If a group of points has a consistent response change direction between five consecutive horizontal points and the spacing between points is 1, then it is merged into a continuous area. The area that meets both the conditions of spatial proximity and channel direction consistency is identified as a response feature segment. Each area needs to record the start and end coordinate indexes, and combine them with the channel number to form a triplet for subsequent recognition operations. All triplet sets that meet the conditions are summarized to obtain a focused feature area set.

[0027] The specific steps of S2 are: S201: Based on the set of focused feature regions, extract the semantic coding content corresponding to the regions in the output sequence, call the sequence index and spatial coordinates of each semantic node, compare the connection order between nodes according to the order of index numbers, and select node pairs with consecutive numbers to obtain a semantic connection structure set; The focus feature region set contains multiple region positions in the image, and each position corresponds to a coding node in the output semantic sequence. For this set, the index number and spatial coordinates of the corresponding node of each region in the semantic sequence are called. When performing number extraction, the one-dimensional serial number in the sequence output tensor is used as the standard, and the two-dimensional coordinate value corresponding to each node in the image space is extracted at the same time, and an index-coordinate comparison table is constructed for subsequent comparison of the connection order. Through this comparison table, all nodes are arranged from small to large according to the serial number, and then the adjacent nodes are sequentially compared according to the arranged number sequence. If a node is numbered n, the next node connected in sequence is n+1. Only node pairs with such a number difference of 1 are retained. If there is a jump number or a number repeat, If the nodes are duplicate, they will be removed from the candidate connections. A simple example is used to illustrate. For example, in a certain image area, the index numbers of the three semantic nodes A, B, and C are 3, 4, and 5, and the corresponding spatial coordinates are (12, 28), (13, 28), and (14, 28). There is number continuity between A and C, and the connection order is consistent, which are valid node pairs. For example, D, E, and F are numbered 5, 6, and 8, and the numbers between F and E are discontinuous. The E to F path should be eliminated. All node pairs that meet the numbering continuity conditions are recorded collectively to form a sequential connection chain of the semantic content in the image. At the same time, node pairs with misplaced numbers due to repeated spatial positions must be excluded to ensure that each group of valid connections has a strict correspondence between serial numbers and spatial positions, and a semantic connection structure set is obtained.

[0028] S202: Based on the direction vector data of each group of connection segments in the semantic connection structure set, the direction angle change value between adjacent path segments is calculated, the path segments with sudden direction changes are filtered out, and the path combinations with continuous direction change amplitude are retained to obtain a set of direction-continuous paths; The calculation formula for the change in the direction angle between adjacent path segments is: ; in, Represents a path segment With path segments The change in the direction angle between Represents a path segment The direction angle, Represents a path segment The direction angle, Represents a path segment The direction vector modulus is Represents a path segment and path segments The modulus of the direction vector difference, represents the exponential decay factor; Assuming a path segment and path segments The starting and ending coordinates are as follows: Path Segment : The starting coordinate is (0, 0) and the end coordinate is (3, 4); Path Segment : The starting coordinates are (3, 4) and the end coordinates are (6, 8); calculate and : ; ; calculate and : ; ; calculate : ; ; The results show that the path segment and path segments The direction angle change value of is zero, which means that the direction angles of the two path segments are the same, and the path segments and path segments The direction of is continuous, which meets the requirements of the direction-continuous path set.

[0029] S203: Based on the spatial coordinates of the nodes in the direction-continuous path set in the output sequence, retrieve the original image position index corresponding to the node, arrange each set of image coordinates according to the connection order, and obtain the regional semantic dependency path sequence; According to the index number of each node, the two-dimensional coordinate parameters in the corresponding output tensor are called in sequence, and the spatial order relationship of all nodes in the path is established. The node index and coordinate binding table are combined to map the coordinate information to the original position in the image frame. By querying the original image position index mapping table, the two-dimensional coordinate value of each output node in the image is located in sequence, and the connection direction in the path is recorded. No jump connection operation is performed to ensure the integrity of the path. Then, for each group of paths, the image coordinates of all nodes are extracted in sequence and sorted according to their connection order. In the example, if there are nodes with numbers 4, 5, 6, and 7, their corresponding image coordinates are (15, 32), (16, 32), (17, 32), and (18, 32), respectively, then the path is represented in the image as a continuous horizontal advancement to the right. If the coordinate order does not match or the node number is repeated, such abnormal path segments need to be removed from the coordinate sequence to ensure that the path sequence direction is single and continuous. Finally, the structural form of the semantic path is represented by image coordinate points and output in groups by path segments, so that the coordinate order in each path is consistent with the connection order, and a regional semantic dependency path sequence is obtained.

[0030] The specific steps of S3 are: S301: Based on the regional semantic dependency path sequence, extract the two-dimensional coordinate information of the nodes in the path within the image frame, number the nodes according to the arrangement order in the path, record the coordinate position data of each pair of adjacent nodes, and arrange the recorded results in sequence to obtain a path node coordinate sequence set; First, for each path, the coordinate position of each node in the image is retrieved one by one, and the corresponding x-axis and y-axis coordinate values ​​are recorded. In the example, the path number P1 contains the node numbers N1, N2, and N3, and their image coordinates are (20, 15), (21, 15), and (22, 15) respectively. This type of coordinate information is obtained by the image position mapping index. Each call needs to be matched and confirmed with the node number to avoid the introduction of wrong coordinates by cross-paths. Then, according to the order of the nodes in the path structure, they are numbered and marked with their order in the path. The numbering is based on the original path order and no reordering operation is required. After numbering, the adjacent numbered nodes are paired in turn in each path to form a coordinate system. Mark pairs and record the position combinations between nodes, such as N1 and N2 form the position pair [(20, 15), (21, 15)], and N2 and N3 form the position pair [(21, 15), (22, 15)]. In this process, no cross-path pairing operation is performed to maintain path independence. Subsequently, each pair of node coordinate data is written into the path position information sequence in numerical order to construct a complete path coordinate point pair record table. Independent record sets are generated for different paths and merged. In the example, P1 and P2 generate node coordinate sequence records for their respective path segments respectively. Finally, the two-dimensional coordinate combination data grouped by path and arranged in order by nodes is output to obtain the path node coordinate sequence set.

[0031] S302: Based on the path node coordinate sequence set, the horizontal displacement value of each pair of adjacent nodes is calculated, the spatial offset direction between the adjacent nodes is identified, the frequency of direction changes in continuous node segments is analyzed, and the coordinate segments with continuous direction trends are marked to obtain the node distribution segments with consistent direction. The calculation formula for the horizontal displacement value of each pair of adjacent nodes is as follows: ; in, Represents the path The horizontal displacement eigenvalue between a node and its adjacent nodes, Representative The horizontal coordinate of each path node, Representatives and The horizontal coordinates of the nodes adjacent to the node, Represents the distance between nodes, Representative The vertical coordinate of the path node, Representatives and The vertical coordinates of the nodes adjacent to the node, Representative The vertical coordinates of the path nodes, Representatives and The vertical coordinates of the nodes adjacent to the node, represents the terrain factor, represents the horizontal coordinate difference, represents the vertical coordinate difference, represents the vertical coordinate difference, represents the error term; Assumptions: The coordinates of path nodes 5 and 7 are as follows: 5th node: rice, rice, rice; 7th node: rice, rice, rice; Set as , Set to 0.001; Calculation process: Calculate the difference in lateral coordinates: rice; Calculate the vertical coordinate difference: rice; Calculate the vertical coordinate difference: rice; Substitute the formula to calculate the horizontal displacement eigenvalue : ; The results show that the horizontal displacement eigenvalue between the fifth and seventh nodes is 0.643 meters. This result is of practical significance in path optimization. By considering the influence of terrain factors, the spatial variation between nodes can be more accurately described. Based on this calculation, it is possible to further analyze the directional changes along the path and mark areas with consistent direction in consecutive node segments, thereby improving the accuracy of path planning and optimization.

[0032] S303: Based on the node distribution segments with consistent directions, select node groups whose adjacent node sequence numbers are continuous, divide the node sequence into consecutive number combinations, and list the node index groups according to the original arrangement order of the path to obtain a semantically guided path set; First, locate all the node numbers in each segment, and use the number sequence to determine whether the adjacent node numbers are sequentially adjacent. For example, if the node numbers in the group are 8, 9, 10, and 11, they can be identified as a continuous node group. If the numbers are 8, 10, and 11, they need to be split into different groups. The number list in the entire node distribution segment needs to be scanned in sequence, and an index table of the sequential relationship between the numbers needs to be established. The nodes with a number difference of 1 are grouped together, and the position where the number jumps is used as the starting point of the new group. After the above number classification operation is completed, the original path order confirmation operation is performed on each group of continuous nodes, that is, the index order of each group of nodes is re-compared to maintain the relative order of the nodes in the original path. The order cannot be confused due to continuous numbers. If There is a path number group of [7, 8, 9], but it is sorted as [8, 9, 7] in the original path. It should be adjusted to [7, 8, 9] and output in sequence, and the index order relationship within the group should be retained. After the node group is established, its index value is added to the new structure list segment by segment. Each node group is treated as a separate sub-path, and only combinations with complete sequence relationships are retained. Node groups with broken or repeated numbers are not included. In the example, the node groups numbered [3, 4, 5] and [7, 8, 9] constitute two groups of path segments, while the node group numbered [3, 5, 6] needs to be split into a single group of [5, 6]. All node groups that meet the order continuity are written into a unified structure data table to form a number grouping list, and finally a semantically guided path set is obtained.

[0033] The specific steps of S4 are: S401: Based on the semantic guidance path set, the sequence of semantic feature values ​​of nodes in the path is analyzed, the continuous changes between the semantic values ​​of adjacent nodes are analyzed, and the connecting segments with different change amplitudes in the sequence are identified to obtain semantic offset paragraph information; First, the semantic feature values ​​of all nodes in each group of paths are extracted. The feature values ​​are derived from the semantic coding values ​​corresponding to the nodes in the output sequence. The feature values ​​corresponding to each node are read and arranged in order according to the path number. For each pair of adjacent nodes in the path, their sequence numbers are called to confirm the node pair combination, and the semantic feature values ​​of the two nodes are taken out for difference to obtain their change values. The positive and negative signs of the numerical results are judged and the change direction is recorded. After the difference operation is completed for all node pairs, the relative differences between the change amplitudes are checked item by item to determine whether the change direction of the adjacent node pairs is continuous. If there is a change segment with a direction reversal or a relatively prominent value span, the starting position of the segment is marked as For the change segment identification point, take the example path number group [2, 3, 4, 5, 6, 7, 8] as an example. If the semantic difference between nodes 2 and 3 is 0.1, 3 to 4 is 0.2, and 4 to 5 is 0.1, but the change between 5 and 6 is 1.3, 6 to 7 is 1.5, and 7 to 8 is 1.2, then 5 to 8 can be considered as a mutation segment. Its starting index 5 and ending index 8 need to be recorded, and its paragraph belonging range needs to be re-marked based on the original path number. This operation needs to cover all path groups and process each segment content separately. Cross-group identification of change differences is not allowed. After processing is completed, the coordinates and number information of all connection segments with mutation characteristics are written to the output list, and finally the semantic shift paragraph information is obtained.

[0034] S402: Based on the semantic offset paragraph information, a sequence comparison is performed on the displacement changes of the path nodes along the change direction, and the jump boundary is divided according to the distribution density of the semantic difference between the nodes to obtain the node jump change mark value; Extract the node numbers and their corresponding coordinate sequence data in each paragraph. Under the premise of keeping the original node arrangement order unchanged, record the coordinate differences between adjacent nodes in the horizontal and vertical directions in the image space, and set the direction identification value to represent the displacement trend. Then, based on the direction of semantic value change between nodes, they are combined and confirmed with their spatial direction. If the continuous nodes maintain a consistent trend in both semantic value difference and spatial displacement direction, this part of the node combination is identified as a coherent segment with a changing trend. Then, for each node pair within the coherent segment of the trend, the cumulative value of the semantic change at each unit position is calculated according to the distribution characteristics of its semantic value difference. A sliding window is formed with every 5 nodes as a group, and its semantic value is counted. The total semantic difference value records the changes in the difference values ​​between windows in order to observe where the jump points with significantly increased semantic differences between adjacent windows appear. For example, the total semantic difference value of nodes 1-5 corresponding to window 1 is 0.8, the total semantic difference value of nodes 2-6 corresponding to window 2 is 0.9, and the total semantic difference value of nodes 3-7 corresponding to window 3 is 1.0. If the value of nodes 4-8 corresponding to window 4 suddenly changes to 2.6, the starting point of the jump interval can be identified as node 4, and the end point can be determined as node 7 by the value recovery stability point in the continuous window. The numbers of nodes 4 to 7 are marked as jump segments, and then a mark value of 1 is set for each node, and the rest are 0. The position index of the jump node in all paths is recorded, and finally the node jump change mark value is obtained.

[0035] S403: Based on the node jump change flag value, the node positions with sudden offset in the path structure are marked with coordinates within the image area and serially numbered. The corresponding image coordinates and numbers are sequentially calibrated according to the order of the nodes in the path to obtain a set of semantic jump node positions. First, the spatial coordinates of each node and the image index number in the original image are read in turn from all the path structures, the coordinate data of the jump node with a mark value of 1 is called, and its position information in the path sequence is extracted as the identification number. Then, the corresponding coordinate position of the jump node in the image frame is matched with its number, and its original arrangement order in the path is recorded to ensure that the image coordinate information of each node corresponds to the number information one by one. Subsequently, under the premise that the node order has not been rearranged, the numbering operation is performed according to the order position of the marked node. That is, if the current jump node is the mth node in the path, its number value is set to m, and its The coordinates are stored in the jump position table. After processing all the jump nodes in the entire path, the number column and the coordinate column are bidirectionally verified to eliminate duplicate records or redundant nodes, and a difference check operation is performed on the integrity and continuity of the number. For example, if there is a jump or missing node number, the empty record that fills the missing item between adjacent numbers is marked as an item with no coordinate value. After the marking is completed, all jump nodes are arranged from small to large according to the number, and a mapping relationship table between the image coordinate sequence and index number corresponding to the jump node is constructed. The table is split and sorted into jump node record items corresponding to multiple paths in the order of the path, and finally a semantic jump node position set is obtained.

[0036] The specific steps of S5 are: S501: Based on the semantic transition node position set, locate the channel region to which each node belongs in the multi-channel response structure, extract the node response amplitude item in the channel direction, and arrange them in sequence according to the channel number order and the node spatial position order to obtain a channel response mapping sequence; Call the row and column index of the node in the channel response feature map, and locate the response value of the corresponding position in each channel according to the index, and then extract the response amplitude of the node position in each channel one by one in combination with the total number of channels. For example, node p is located in the xth row and yth column in the image frame, and the response amplitude of the position (x, y) is read channel by channel in channels T0 to Tn, and recorded as a set of response amplitude vectors. This process is executed in a loop on all jump nodes. Subsequently, the channel response amplitude vectors extracted from each node are arranged according to their serial numbers in the path structure, that is, the order in which the nodes appear in the original path is established. Index, if the node number is 1, 2, 3, then the corresponding response vectors are also arranged in sequence. Then, within each group of vectors, the element order is arranged in the channel number order T0→T1→T2... to ensure the uniformity of the data structure across nodes and channels. Finally, a two-layer sequence structure is constructed, one for the node sequence and the other for the channel sequence within each node. Under this structure, the combination of all node response vectors is completed, and a complete permutation table for the joint mapping of channel number and node space order is constructed. The table can be represented as a two-dimensional matrix, with the column dimension being the channel number and the row dimension being the node number. Finally, the channel response mapping sequence is obtained.

[0037] S502: Based on the channel response mapping sequence, the arrangement order of the continuous nodes is compared with the boundary information of the channel active area, and the position group where the distribution continuity is interrupted in the spatial connection structure within the channel is identified to obtain the characteristic distribution fracture information; First, in the sequence, the arrangement order of the nodes is used as the reference frame, and the response position relationship between the node pairs is extracted channel by channel. The coordinate point pairs with node numbers n and n+1 are taken as a group, and the two-dimensional spatial position corresponding to the response amplitude of the group in each channel is extracted. After the spatial coordinate range of the corresponding image area is fixed, the active area boundary inside the channel is used as the comparison benchmark to read whether the pair of node coordinates are within the same area boundary range. If the channel responses of two consecutive points are located in different boundary areas, it is recorded as a boundary fracture event. According to this standard, the boundary cross-region comparison operation of all adjacent node pairs in the node sequence is carried out in sequence. Whenever a cross-border behavior with discontinuous position occurs, the correspondence between the node number and the channel number is recorded to form an initial record set of fracture marks. In order to improve the operation accuracy, the channel number needs to be controlled in the main sequence during the recording process, that is, Starting from the channel with the smallest number, the horizontal determination of the response position is performed in sequence in an ascending order of numbers. This ensures that the spatial continuity of each node group in the channel dimension is not missed. Subsequently, all node pairs with breaks are respectively classified into channel grouping containers. Each container indicates the node position number where the spatial connection is interrupted. Combined with the coordinate jump behavior between the spatial positions of the preceding and succeeding nodes, the jump area span and direction consistency can be used to verify again. For example, when the spatial offset distance between the response positions of nodes 8 and 9 in channel 3 is greater than 4 pixel units, and the two points fall into two non-adjacent response distribution blocks, the node pair is identified as a structural connection interruption group. Finally, the position combinations of all node pairs with interruptions in all channels are screened according to the above process and merged into a complete list to obtain the characteristic distribution break information.

[0038] S503: calling feature distribution fracture information, locating the corresponding position of the jump node segment in the image area, extending the edge linear structure direction according to the connection boundary change direction, and obtaining a semantic response area distribution structure diagram; First, obtain the node index of the path segment where the jump node is located, extract the node group with the corresponding index in the path structure, call the two-dimensional coordinate value in its image area according to the node index, and record the coordinate set corresponding to all jump node segments as the basis for image mapping positioning. Then, along the connection direction of each pair of consecutive node pairs in the set, calculate the coordinate difference of each pair of nodes in the X-axis direction and the Y-axis direction, and determine the direction of the connection boundary according to the size and positive and negative of the difference. For example, if the X-direction difference of adjacent nodes is positive and the Y-direction difference is zero, the direction of the connection segment points to the right side of the image. If the X-direction and Y-direction differences are both positive, the direction tends to the lower right side of the image. After the connection direction is clear, a set of auxiliary point sets is constructed by calling the extension line of the direction vector to simulate the edge extension trajectory. The position of each auxiliary point needs to be based on the previous The node vector direction is advanced by a coordinate increment of a unit distance, and the point is appended to the extension path list in a linear arrangement in the image coordinate space. During the execution process, the image boundary restriction conditions are combined to limit the extension length to not exceed the image size boundary. At the same time, points with coordinate values ​​less than 0 or exceeding the maximum length and width of the image are skipped. Finally, for each jump node fragment, its position path after projection in the image area is obtained, and the edge extension trajectory of each group of paths is organized in the connection order. By drawing all trajectory segments in sequence in the image plane, a linear structure set covering the jump fragment is obtained. The starting and ending points of each line segment in the set correspond to the actual image coordinate position. By connecting the line segments, a complete semantic response edge distribution is formed. Combined with the original channel response layer of the image, the structure is superimposed and displayed to obtain a semantic response area distribution structure diagram.

[0039] See also Figure 2 , a processing system based on a large computer vision model, comprising: The feature focus extraction module obtains the feature map of the input image frame through the multi-layer feature extraction structure. Based on the difference in response amplitude in each channel direction, it identifies the areas with continuous response at the same position, aligns the positions of the areas at each level, and marks the areas with consistent change characteristics across layers to obtain a set of focused feature areas. The semantic path construction module extracts the semantic information content of the corresponding region in the output sequence based on the set of focused feature regions. It calls the connection direction data and semantic sequence data, analyzes the consistency of the path direction between semantic nodes, selects the path information with consistent connection direction and continuous semantic relationship, and marks the corresponding position on the original image to obtain the regional semantic dependency path sequence. The path sequence classification module extracts the spatial coordinate order of the nodes in the path based on the regional semantic dependency path sequence, compares the order consistency of the arrangement directions of adjacent coordinate points, and classifies the node sequences with prominent continuity according to the position advancement relationship to obtain the semantically guided path set; The transition node positioning module tracks the change in the semantic feature value of each node in the path between consecutive positions based on the semantic guidance path set. It then tracks the change position according to the offset trend and marks the image coordinates and index number of the transition point. This generates a set of semantic transition node positions and passes it to the response area recognition module. The response area identification module extracts the characteristic content of each node in the channel response structure based on the semantic jump node position set, compares the position arrangement between adjacent nodes with the distribution results of active areas in the channel, identifies the position segments where the characteristic response mutates, and delineates them as independent response segments to obtain the semantic response area distribution structure diagram.

[0040] 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 modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A processing method based on a large computer vision model, characterized in that: The following steps are involved: S1: Obtain the feature map of the input image frame through the multi-layer feature extraction structure, extract the continuous response area of ​​the same position in the channel, align the cross-layer area positions, filter the areas with consistent change directions, and obtain the set of focused feature areas; S2: Based on the set of focused feature regions, extract the semantic information nodes of the regions in the output sequence, analyze the node connection direction and semantic order, screen the path clues with consistent directions, and map the path positions to the original image regions to obtain the regional semantic dependency path sequence; S3: Based on the regional semantic dependency path sequence, extract the spatial coordinates of the nodes in the path, analyze the consistency of the arrangement directions of adjacent coordinates, group the nodes with continuous advancement directions into one group, and obtain a semantic guidance path set; S4: Based on the semantic guidance path set, tracking the change trend between the semantic values ​​of the nodes in the path, marking the jump nodes with position offsets, recording the corresponding coordinates, and obtaining a semantic jump node position set; S5: Based on the semantic jump node position set, extract the local features of the nodes in the channel response structure, analyze the spatial arrangement and activation area differences between adjacent nodes, delineate the mutation segments, and obtain the semantic response area distribution structure diagram.

2. The processing method based on a computer vision large model according to claim 1, characterized in that: The focused feature area set includes continuous response areas, inter-layer alignment information, and cross-layer consistency labels. The area semantic dependency path sequence includes semantic information points, connection order, path direction, and image position annotations. The semantic guidance path set includes spatial coordinate order, direction consistency sequence, and continuity node category. The semantic jump node position set includes jump node image coordinates, index numbers, and offset trend points. The semantic response area distribution structure diagram includes feature response mutation positions, channel activity distribution differences, and independent response fragments.

3. The processing method based on a computer vision large model according to claim 1, characterized in that: The specific steps of S1 are: S101: Obtain multiple sets of feature maps of the input image frame through a multi-layer feature extraction structure, call the response amplitude value of the same position in the channel, compare the response amplitude change of the coordinate point according to the channel number, and select the coordinate point area with the same response amplitude change direction on the continuous channel to obtain the response continuous point group; S102: Based on the response continuation point group, call the position index value in each layer feature map, compare the position correspondence relationship of the same point between the differentiated layers, filter out the points with discontinuous cross-layer coordinate correlation relationship, and obtain the layer-aligned position group; S103: Based on the response amplitude of each position in the hierarchical alignment position group in the channel direction, and comparing the channel response direction change, screen the point areas whose response change trends remain consistent between differentiated channels, identify the continuous point distribution segments with change characteristics, and obtain a set of focused feature areas.

4. The processing method based on a computer vision large model according to claim 1, characterized in that: The specific steps of S2 are: S201: Based on the set of focused feature regions, extract the semantic coding content corresponding to the regions in the output sequence, call the sequence index and spatial coordinates of each semantic node, compare the connection order between nodes according to the order of index numbers, and select node pairs with consecutive numbers to obtain a semantic connection structure set; S202: Calculating the direction angle change between adjacent path segments based on the direction vector data of each group of connection segments in the semantic connection structure set, filtering out path segments with sudden direction changes, and retaining path combinations with continuous direction change amplitudes to obtain a set of direction-continuous paths; S203: Based on the spatial coordinates of the nodes in the direction-continuous path set in the output sequence, retrieve the original image position index corresponding to the node, arrange each group of image coordinates according to the connection order, and obtain the regional semantic dependency path sequence.

5. The processing method based on a computer vision large model according to claim 4, characterized in that: The calculation formula of the change value of the direction angle between adjacent path segments is specifically: ； in, Represents a path segment With path segments The change in the direction angle between Represents a path segment The direction angle, Represents a path segment The direction angle, Represents a path segment The direction vector modulus is Represents a path segment and path segments The modulus of the direction vector difference, represents the exponential decay factor.

6. The processing method based on a computer vision large model according to claim 1, characterized in that: The specific steps of S3 are: S301: Based on the region semantic dependency path sequence, extract the two-dimensional coordinate information of the nodes in the path within the image frame, number the nodes according to the arrangement order in the path, record the coordinate position data of each pair of adjacent nodes, and arrange the recorded results in sequence to obtain a path node coordinate sequence set; S302: Based on the path node coordinate sequence set, calculate the horizontal displacement value of each pair of adjacent nodes, identify the spatial offset direction between adjacent nodes, analyze the frequency of direction changes in continuous node segments, mark the coordinate segments with continuous direction trends, and obtain node distribution segments with consistent directions; S303: Based on the node distribution segments with consistent directions, filter the node groups whose adjacent node sequence numbers are continuous, divide the node sequence into continuous number combinations, and list the node index groups according to the original arrangement order of the path to obtain a semantically guided path set.

7. The processing method based on a computer vision large model according to claim 6, characterized in that: The calculation formula for the horizontal displacement value of each pair of adjacent nodes is specifically: ； in, Represents the path The horizontal displacement eigenvalue between a node and its adjacent nodes, Representative The horizontal coordinate of each path node, Representatives and The horizontal coordinates of the nodes adjacent to the node, Represents the distance between nodes, Representative The vertical coordinate of the path node, Representatives and The vertical coordinates of the nodes adjacent to the node, Representative The vertical coordinates of the path nodes, Representatives and The vertical coordinates of the nodes adjacent to the node, represents the terrain factor, represents the horizontal coordinate difference, represents the vertical coordinate difference, represents the vertical coordinate difference, represents the error term.

8. The processing method based on a computer vision large model according to claim 1, characterized in that: The specific steps of S4 are: S401: Based on the semantic guidance path set, analyzing the sequence of semantic feature values ​​of nodes in the path, continuously changing the semantic values ​​of adjacent nodes, identifying connected segments with different change amplitudes in the sequence, and obtaining semantic offset paragraph information; S402: Based on the semantic offset paragraph information, perform sequence comparison on the displacement changes of the path nodes along the change direction, divide the jump boundary according to the distribution density of the semantic difference between the nodes, and obtain the node jump change mark value; S403: Based on the node jump change mark value, the node positions of the offset mutation in the path structure are subjected to coordinate marking and serial numbering in the image area, and the corresponding image coordinates and numbers are calibrated in sequence according to the sequential arrangement of the nodes in the path to obtain a semantic jump node position set.

9. The processing method based on a computer vision large model according to claim 1, characterized in that: The specific steps of S5 are: S501: Based on the semantic transition node position set, locate the channel region to which each node belongs in the multi-channel response structure, extract the node response amplitude item in the channel direction, and arrange them in sequence according to the channel number order and the node spatial position order to obtain a channel response mapping sequence; S502: Based on the channel response mapping sequence, compare the arrangement order of continuous nodes with the boundary information of the channel active area, identify the position group where the distribution continuity is interrupted in the spatial connection structure within the channel, and obtain characteristic distribution fracture information; S503: calling the feature distribution fracture information, locating the corresponding position of the jump node segment in the image area, extending the edge linear structure direction according to the connection boundary change direction, and obtaining the semantic response area distribution structure diagram.

10. A processing system based on a large computer vision model, characterized in that: The system is used to implement the processing method based on a computer vision large model according to any one of claims 1 to 9, and the system includes: The feature focus extraction module obtains the feature map of the input image frame through the multi-layer feature extraction structure. Based on the difference in response amplitude in each channel direction, it identifies the areas with continuous response at the same position, aligns the positions of the areas at each level, and marks the areas with consistent change characteristics across layers to obtain a set of focused feature areas. The semantic path construction module extracts the semantic information content of the corresponding region in the output sequence based on the set of focused feature regions. It calls the connection direction data and semantic sequence data, analyzes the consistency of the path direction between semantic nodes, selects the path information with consistent connection direction and continuous semantic relationship, and marks the corresponding position on the original image to obtain the regional semantic dependency path sequence. The path sequence classification module extracts the spatial coordinate order of the nodes in the path based on the regional semantic dependency path sequence, compares the order consistency of the arrangement directions of adjacent coordinate points, and classifies the node sequences with prominent continuity according to the position advancement relationship to obtain the semantically guided path set; The transition node positioning module tracks the change in the semantic feature value of each node in the path between consecutive positions based on the semantic guidance path set. It then tracks the change position according to the offset trend and marks the image coordinates and index number of the transition point. This generates a set of semantic transition node positions and passes it to the response area recognition module. The response area identification module extracts the characteristic content of each node in the channel response structure based on the semantic jump node position set, compares the position arrangement between adjacent nodes with the distribution results of active areas in the channel, identifies the position segments where the characteristic response mutates, and delineates them as independent response segments to obtain the semantic response area distribution structure diagram.

Images