A terrain monitoring method

Abstract

The present invention relates to a terrain monitoring method, comprising the following steps: deploying targets at monitoring points and acquiring image samples; constructing a target detection model to detect the image samples, and outputting a first monitoring signal when the target detection model fails to detect the target or when the target detection model detects that the target has deviated from its original position; the target detection model comprises a plurality of nodes with correlation coefficients as activation functions, the nodes comprising a core node for extracting pixel feature units of an input image, and multi-level segment nodes corresponding to pixel features of different scales of the image sample, formed based on the spatial relationship of the pixel feature units, the core nodes and the segment nodes forming a tree structure that converges upward step by step from the core nodes, and outputting the target detection position through the segment nodes at the root. The present invention can accurately and efficiently monitor target terrain and promptly issue terrain change warning signals.

Description

Technical Field

[0001] The present invention relates to the technical field of terrain monitoring, and in particular to a terrain monitoring method. Background Art

[0002] Geological disasters such as landslides, mudslides, and dam breaches require a system with rapid response and high reliability. To increase the basis for judgment and enable maintenance personnel to make more accurate disaster prevention responses, the system should incorporate more terrain beyond the monitoring points. The diverse topography and complex environment often make system deployment and maintenance difficult, and changes in the vegetation covering the terrain also pose greater challenges.

[0003] The existing geological monitoring system mainly has the following implementation methods:

[0004] 1. Using tension wires and stress sensors, etc., to detect stress changes and then assume that geological changes have occurred;

[0005] 2. Use a water level sensor or rainfall sensor to sound an alarm when the humidity exceeds the warning value;

[0006] 3. Use multiple sensors to model soil characteristics and then estimate disaster risk based on the model and detection values;

[0007] 4. Use convolutional neural networks to monitor terrain or landmarks.

[0008] Among them, the first three methods often require the deployment of multiple sensors, which are difficult and costly to deploy and maintain; modeling requires a large number of experiments, and fine-tuning is required for mountains with different components; although the fourth method has simple sensors, network training has high requirements for data, computing, and time resources; for neural network systems that directly monitor terrain changes, their accuracy is easily affected by conditions such as vegetation changes, light and shadows; for neural network systems that only monitor targets, their monitoring range is small and their accuracy is easily affected by the selection of target marking locations.

[0009] Therefore, there is a need for a geological monitoring system that can stably and reliably conduct comprehensive and continuous monitoring of monitoring points and surrounding terrain, take into account multiple working modes, communication modes and power supply modes, has stronger adaptability to complex natural environments and can significantly reduce the time and cost of deployment, maintenance and system construction. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a terrain monitoring method that can accurately and efficiently monitor the target terrain and issue terrain change warning signals in a timely manner.

[0011] The technical solution adopted by the present invention to solve the technical problem is: to provide a terrain monitoring method, comprising the following steps:

[0012] Deploy targets at monitoring points and acquire image samples;

[0013] Constructing a target detection model to detect the image sample, and outputting a first monitoring signal when the target detection model fails to detect the target or when the target detection model detects that the target deviates from an original position;

[0014] The target detection model includes multiple nodes with correlation coefficients as activation functions. The nodes include core nodes used to extract pixel feature units of the input image, and multi-level segment nodes corresponding to pixel features of different scales of the input image formed based on the spatial relationship of the pixel feature units. The core nodes and the segment nodes constitute a tree structure that converges upward step by step from the core nodes, and outputs the target detection position through the segment node at the root.

[0015] Furthermore, the target detection model is trained by the following method:

[0016] Obtaining an initial image sample, and generating an object detection model based on the initial image sample;

[0017] Acquire a training image sample and a valid pixel area of ​​the training image sample;

[0018] Detecting the training image samples using the target detection model to obtain target detection positions;

[0019] Based on the structure of the target detection model, backtracking to obtain the activated pixel area of ​​the target detection position;

[0020] The network structure of the target detection model is optimized according to the similarity between the activated pixel area and the effective pixel area.

[0021] Furthermore, the optimizing the network structure of the target detection model according to the similarity between the activated pixel region and the effective pixel region includes:

[0022] Calculating the pixel difference between the activated pixel area and the effective pixel area;

[0023] Dividing a plurality of false detection areas according to the pixel difference;

[0024] Backtracking to obtain a first optimized node, where the first optimized node is the highest-level node on the branch where the first optimized node is located and is completely covered by the false detection area;

[0025] A second type of shadow node is embedded between the first optimization node and its parent node to conditionally activate the pixel features corresponding to the first optimization node.

[0026] Furthermore, the second type shadow node with a positive condition embedded between the first optimized node and its parent node conditionally activates the pixel feature corresponding to the first optimized node, including:

[0027] Determine whether there is a positive condition trigger node, where the positive condition trigger node is an inactivated node at the same level as the first optimization node and adjacent to the corresponding pixel feature;

[0028] If the positive condition trigger node exists, construct a positive condition second type shadow node, and use the first optimization node and the positive condition trigger node under the same parent node as child nodes of the positive condition second type shadow node. When the positive condition second type shadow node is anchored by the positive condition trigger node and the positive condition trigger node is activated, the positive condition second type shadow node is activated.

[0029] If the positive condition trigger node does not exist, a second type of shadow node with a negative condition is constructed to conditionally activate the pixel feature corresponding to the first optimization node.

[0030] Furthermore, the second type of shadow node for constructing the negative condition includes:

[0031] Acquire a number of activated nodes located at the junction of the false detection area and the valid pixel area as a shadow set;

[0032] Constructing a false detection branch structure corresponding to pixel features of different scales in the false detection area, and using the top node of the false detection branch structure as a negative condition trigger node;

[0033] A second type of shadow node of the negative condition is constructed with the shadow set and the false detection branch structure as child nodes.

[0034] Furthermore, it is set that when the nodes in the shadow set and the negative trigger node are both activated and anchored to each other, the second type of shadow node of the negative condition places the negative trigger node in an inactive state, so that the false detection branch structure is not detected by the target detection model.

[0035] Furthermore, after the step of calculating the pixel difference between the activated pixel area and the effective pixel area, the method further includes:

[0036] Dividing a plurality of missed detection areas according to the pixel difference;

[0037] Backtracking to obtain a second optimized node, where the second optimized node is the highest-level node on the branch where it is located and is completely covered by the missed-detection area;

[0038] The second optimized node is replaced by a first-type shadow node, wherein the child nodes of the first-type shadow node include the second optimized node and its replacement node, and when any one of the second optimized node and its replacement node is activated, the first-type shadow node is activated.

[0039] Furthermore, the replacement node is one of the activated nodes at the same level as the second optimization node and with similar corresponding pixel features, and the center coordinates of the first shadow-like node, the replacement node and the second optimization node are all set to be the same.

[0040] Furthermore, after the step of dividing a plurality of missed detection areas according to the pixel difference, the method further includes:

[0041] Determine the missing area of ​​the missed detection area in the target detection model;

[0042] Constructing a missed detection branch structure corresponding to pixel features of different scales in the missing area;

[0043] Selecting an activated node at the same level as the top node of the missed-detection branch structure as a joint triggering node;

[0044] The associated trigger node and the top node of the missed-detection branch structure are used as child nodes of the third-type shadow node, and when the associated trigger node is activated and the child nodes of the third-type shadow node can all be anchored, the third-type shadow node is activated;

[0045] The third type of shadow nodes are added to the target detection model.

[0046] Beneficial effects

[0047] Due to the adoption of the above-mentioned technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention sets targets and auxiliary identification points on the target terrain, monitors the geological changes of the target marked location with high priority, and takes other terrains into account with lower priority. Compared with the system that directly monitors the terrain changes in the entire picture, it can better eliminate the influence of vegetation and the like on the detection results; the present invention adopts the correlation coefficient as the feedforward function and activation function to construct the target detection model, and can learn the target features and perform identification and positioning through the single sample or small sample recognition technology without pre-training, without the need for manual intervention in hyperparameter design and autonomous construction of the neural network structure; the present invention utilizes the strong traceability of the target detection model, and updates the network structure by comparing the similarity between the activated pixel area of ​​the model and the reference effective pixel area of ​​the training sample, without the need to adopt the gradient descent method, and the training is simple and efficient; the present invention also designs three types of shadow nodes with different logical structures, and uses these three types of shadow nodes to optimize the model to make the prediction results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a schematic diagram of the deployment of an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a target detection model according to an embodiment of the present invention;

[0050] Figure 3 is a flow chart of target detection model training according to an embodiment of the present invention;

[0051] Figure 4 is a schematic diagram of FP corresponding to a branch node according to an embodiment of the present invention;

[0052] Figure 5 Schematic diagram of a method for constructing a second type of shadow node with positive conditions according to an embodiment of the present invention;

[0053] Figure 6 Schematic diagram of a method for constructing a second type of shadow node with negative conditions according to an embodiment of the present invention;

[0054] Figure 7 2 is a schematic diagram of a method for constructing a first type of shadow node according to an embodiment of the present invention;

[0055] Figure 8 2 is a schematic diagram of a method for constructing a third type of shadow node according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall equally within the scope limited by the appended claims of the application.

[0057] The embodiment of the present invention relates to an intelligent terrain monitoring method based on an improved neural network, comprising the following steps:

[0058] like Figure 1 As shown, targets are set on the monitored terrain (such as hillsides and river banks), and on-site processors are set up on ridges or other relatively fixed locations. Cameras monitor the targets and their surrounding natural terrain landmarks, and the central control system is set up in a place that is easy to guard.

[0059] When the surveillance system goes live, live video feeds containing targets can be used to directly train the intelligent software. During rollout and maintenance, natural terrain landmarks or other image features can be manually added as auxiliary markers for early warning. When adding natural terrain landmarks, it's best to select areas without significant vegetation cover, such as exposed rock faces or flood control markings, to minimize interference and improve reliability. This monitoring approach, which prioritizes targets and supplements natural terrain landmarks, provides more complete coverage of the monitored terrain and minimizes unavoidable interference from vegetation, water flow, and other factors.

[0060] The system intermittently samples the target and its surroundings, with sampling intervals adjustable based on monitoring requirements and power and communication conditions. The system can be configured for either front-end or back-end processing, with the former enabling on-site processors to process and compare images and only sending processed data to the central control system. The latter allows the on-site processors to send the collected images and raw data to the central control system, which then performs image processing and comparison.

[0061] If the target is detected as deviated or lost, indicating a change in the natural terrain, the central control system issues a warning signal. If the target in the image shifts from its original position or disappears, indicating a relative displacement between the landmark and the camera, the central control system issues an alarm signal. Upon the issuance of these two signals, the central control system automatically retrieves the abnormal image for manual review. If the central control system times out without receiving information from the on-site processor, this indicates an unexpected loss of the on-site processor signal, indicating damage. This indicates a change in the terrain, posing a risk of geological disasters such as landslides, collapses, and mudslides, and the central control system issues a loss alarm.

[0062] To adapt to complex environments and better meet the needs of field use, the system's front-end processor includes built-in wired and mobile communication modules, a solar module, and a battery. This allows for continued data transmission even when power and communication quality are poor, or allows for fully wireless operation through scheduled manual inspections and battery replacement. Front-end or back-end processing modes can be set based on energy and communication conditions, and sampling intervals can be set based on monitoring security levels, ensuring consistent monitoring while conserving energy and extending service life.

[0063] The system uses an improved neural network to build a target detection model. Its feedforward and activation functions are based on correlation coefficients, eliminating the need for pre-training, gradient descent, or repeated manual adjustments to the network structure. The network automatically adjusts its structure based on the complexity of the target, minimizing resource consumption. Training and calibration are completed using only the first few frames, allowing real-time deployment using live data.

[0064] like Figure 2As shown, the object detection model of this embodiment includes multiple nodes using correlation coefficients as activation and feedforward functions. Specifically, it includes a core node for extracting pixel feature units from the input image, and a multi-level segment node, corresponding to pixel features at different scales of the image sample, formed based on the spatial relationship between these pixel feature units. The core nodes and segment nodes ultimately form a tree structure that converges from the core node upward, and the target detection location is output through the segment node at the root.

[0065] These pixel feature units correspond to different activated areas in the input image, and there may be overlap between areas, and their spatial positions are represented by central coordinates. Each segment node in the tree structure is the sum of the features of all its child nodes, and corresponds to the image area contained in all child nodes, and its spatial position is also represented by central coordinates. When the image to be detected is input into the model, each node is activated and anchored step by step from low to high according to the calculation results of the correlation coefficient (when the correlation coefficient with the corresponding image area is greater than the threshold, the node is activated, and when the correlation coefficient with the adjacent node is greater than the threshold, the node is anchored). Finally, the root node outputs the predicted coordinates.

[0066] The above target detection model is trained by the following method, such as Figure 3 As shown:

[0067] The first step is to obtain the training image and the valid pixel region of the training image. The valid pixel region class can be manually delineated or completed by a program.

[0068] The second step is to use the object detection model to detect the training image and determine the target detection position. If the target exists and has not deviated from its original position, the valid pixel area is updated based on the target detection position. When determining whether the target has deviated from its original position, a threshold is set to prevent false triggering of the warning signal. The target is considered to have deviated only when the target offset exceeds the threshold. Therefore, in order to obtain more accurate training results, the valid pixel area needs to be updated.

[0069] The third step is to trace back the activated pixel region at the target detection location based on the structure of the target detection model. Because the model's activation function and feedforward function are both constructed based on correlation coefficients, network parameters are directly generated instead of using a gradient descent algorithm. This results in a highly unified network structure and features, allowing the nodes corresponding to each pixel in the target image to be traced back to the corresponding activated pixels forming the activated pixel region.

[0070] The fourth step is to optimize the network structure of the target detection model based on the similarity between the activated pixel area and the effective pixel area. The specific steps are:

[0071] A0 calculates the pixel difference between the activated pixel area and the reference valid pixel area.

[0072] A1 divides several missed detection areas and false detection areas according to pixel differences, where the missed detection area is the area included in the valid pixel area but not in the activated pixel area, and the false detection area is the area not included in the valid pixel area but included in the activated pixel area.

[0073] A2 adjusts the network structure through shadow nodes for missed detection areas and false detection areas.

[0074] The following further explains the application of shadow nodes in optimizing model structure in conjunction with the specific training process.

[0075] To enrich node functionality and enable more complex functions, this design introduces three types of shadow nodes to handle logical relationships between features. These are: 1. Shadow nodes for OR logic; 2. Shadow nodes for conditional activation; and 3. Shadow nodes for joint activation. Shadow nodes are not counted in the actual number of layers and are not included in the traceback depth.

[0076] The first type is shadow nodes used for OR logic

[0077] This type of shadow node implements the logical relationship of A or B, where A and B are nodes on the same layer. This type of shadow node allows nodes at the same layer and center to replace each other. When pre-adding prior experience or performing small sample training, this type of shadow node can process multiple features as the same feature, reducing the network's sensitivity to changes in a particular feature. This type of shadow node requires the centers of nodes A and B to be aligned. The recorded center is also the center of the shadow node's child nodes (however, shadow nodes are not counted in the hierarchy, so A and B remain nodes on the same layer after adding the shadow node). When the shadow node is activated by its child nodes, the logical relationship of A or B is implemented.

[0078] For example, in the Mnist task, we know that the digit "1" has three angles: " / ," "|," and "\." Normally, these are treated as three different angles. However, for handwritten digits, all three angles can represent "1," so by adding shadow nodes, they can be treated as the same angle. If these pre-added feature nodes representing the digit "1" at different angles are activated on the build set image, they will be treated as normal nodes and combined into higher-level nodes.

[0079] The second type is conditionally activated shadow nodes.

[0080] This type of shadow node involves multiple child nodes. Its activation and anchoring process is similar to that of segment nodes. It can be divided into two types, which are activated under positive and negative conditions respectively.

[0081] Under positive conditions, this type of shadow node requires a positive trigger node A. When all correlation coefficients including trigger node A exceed the threshold, the shadow node is activated, implementing the A and Collection logic. Collection represents a feature activated by child nodes other than trigger node A. That is, while other child nodes can activate the shadow node, it must meet the requirements of A to achieve activation.

[0082] Under the negative condition, a trigger node A' is also required, and Collection is the trigger of point A'. Only when both the trigger node A' and Collection are activated, and the trigger node A' can be anchored by the collection (that is, the correlation coefficient between the trigger node A' and the nodes in the collection is greater than the threshold), the shadow node is activated, thereby suppressing the trigger node A' and making it output an inactive state, so that the FP area corresponding to the trigger node A' will not be detected.

[0083] The third type: shadow nodes activated together

[0084] This type of shadow node also requires a trigger node A' and is activated simultaneously with it by default. That is, if the feature corresponding to the shadow node is activated and A' is also activated, it is directly added to the activation area. The center of this shadow node is recorded as the center of the trigger node A'. If the trigger node A' is activated and all the child nodes of the shadow node are anchored (that is, the relevant calculations of other nodes excluding the A' node are greater than the threshold), the shadow node is activated and its area is increased in the output activation pixels.

[0085] The small sample training process in this implementation essentially edits the network structure to enable it to handle both false-negative and false-positive pixel detection results. This implementation is limited to optimizing the network structure when the terminal node is activated. When the terminal node is inactivated, a new network structure can be created to connect to the terminal node of the same classification, which will not be discussed further.

[0086] First, the original network is run in test mode on images from the training set. Using a minimal backtracking method, the activation features are traced back to determine the actual activated regions. The activated regions are compared with the valid pixel mask of the test image to generate a false positive (FP) region map (i.e., falsely detected regions) and a false negative (FN) region map (i.e., missed regions). While existing neural network errors are typically expressed as a single value, the error in this design is expressed as two pixel error maps. Subsequent training is based on these two maps.

[0087] 1) Training on FP error map:

[0088] During the computation on the training samples, all activated segment nodes are recorded, and the inactive child nodes of the activated segment nodes are also marked. Based on these marked nodes, the network activation structure is traced back to find the branches related to the FP region. The branches are then traced down until their coverage area falls completely within the FP region. The top node of this branch is then used as the first optimized node.

[0089] If the nodes that share the same upper-level node with the first optimized node also fall within the FP region, these nodes are combined. If the number reaches the minimum number of child nodes for a segment node, they are packaged and await further processing. If the number is insufficient, connected activated nodes are randomly selected from adjacent activated nodes at the same level, using the same selection rules as for segment node construction. The selected nodes are then packaged with the original branch vertex.

[0090] Using estimation mode, trace the network structure to find inactive nodes at the same level as the branch vertex and mark their estimated range. Use the estimated range to filter inactive nodes that are connected to the vertex's actual range. Randomly select one of the inactive nodes that meet the requirements as trigger node A. Use trigger node A and the packaged nodes to generate the second type of positive conditional shadow nodes.

[0091] If there is no qualified trigger node A among the activated nodes of the same layer in the original network, several activated nodes located between the FP area and the valid pixel area are selected as the collection, and a new segment node is constructed as the trigger node A'. The segment node and its child nodes correspond to the FP area, and the trigger node A' and its child nodes are used to create the second type of negative conditional shadow node.

[0092] If the number of child nodes below the upper node of the original branch vertex that do not fall within the FP area exceeds the minimum number of child nodes of the segment node - 1 (minus the number of shadow nodes to be added), the branch vertex is deleted, the generated shadow node is added to the upper node, and then the registration parameters of the segment node are updated. If the number of remaining child nodes is insufficient, the activated nodes connected to the remaining child nodes are found according to the rules for finding adjacent nodes when generating segment nodes, and these activated nodes are used to replace the branch vertex and join the upper node, and the shadow node is also added, and finally the parameters of the upper segment node are updated. When updating the parameters, the relative position of the center of the upper segment node is kept unchanged, and the deviation value between the position of the newly added child node and the center of the segment node is recalculated. Since the addition of shadow nodes occupies the number of network layers, the number of layers of all nodes needs to be updated after adding them.

[0093] like Figure 4As shown, the green area is the FP region, the blue hexagon represents the feature represented by a segment node, regions 1 to 4 are the child nodes of the segment node, and 2-1 to 2-4 are the child nodes of child node 2. In the figure, the region activated by node 4 is completely within the FP region, meeting the requirement and requiring no further tracing. However, nodes 1, 2, and 3 are only partially within this region, so further tracing back to their child nodes is necessary. Taking child node 2 as an example, further tracing back reveals that 2-2, 2-3, and 2-4 fall within the region, meeting the requirement. Similarly, nodes 1 and 3 can be traced back to their next-level child nodes. This process is repeated until the core node is reached.

[0094] Figure 5 The training process for the FP region is shown. Nodes 12 and 13 are completely in the FP region, and node 5 is its upper node. An unactivated node N1 adjacent to nodes 12 and 13 in an estimated region is selected as a trigger node. If a segment node has too few child nodes, activated adjacent nodes are added. Nodes N1, 12, and 13 are used to generate a second-class positive conditional shadow node 22. Nodes 12 and 13 are deleted from node 5 and shadow node 22 is added. If the number of child nodes of node 5 is less than the minimum requirement, activated connected nodes are added to it. Finally, the network structure is updated.

[0095] like Figure 6 As shown in the figure, if no qualified trigger node A is found among the activated nodes in the same layer of the original network, several activated nodes located between the FP region and the valid pixel region are selected as a shadow collection. Then, a new segment node is constructed as the trigger node A', and the child nodes of this segment node correspond to the FP region. When the nodes in the collection and the trigger node A' are both activated, and the trigger node A' is anchored to the nodes in the collection, the shadow node forces the trigger node A' to output an inactive state, so that the corresponding FP region will not be detected by the model.

[0096] 2) Training for FN error

[0097] Addressing FN errors essentially involves incorporating the substructure representing that region into the network model. This can occur in two situations: 1. The feature node and its branches that should be activated at that location are not; 2. The region is missing in the original network. These two situations are handled separately.

[0098] For the first case, the inactivated nodes are replaced by the following method.

[0099] In estimation mode, trace back the feature branches that fall completely within the FN region to the top node of the branch that falls completely within the region. Then, among all activated segment nodes, search for nodes on the same layer that have a similar position and coverage area to the inactivated branch's top node. Nodes with a correlation coefficient greater than 0.8 are considered interchangeable.

[0100] Based on the activated nodes and the original inactivated nodes, the first type of shadow nodes are constructed, and the shadow nodes are used to replace the original branch vertices and registered in the upper segment nodes of the original vertices. Finally, the number of layers of the entire network is updated.

[0101] like Figure 7 As shown, points 19-21 are not activated, while nodes 22-24, which are close in location and coverage area, are activated here. Then, a type of shadow node is added to replace them respectively, and the network is updated after connecting to the upper node.

[0102] For the second case, a new FN area branch is added using the following method.

[0103] First, a subnetwork is constructed on the FN region according to the network construction method. The core nodes of the original network can be reused during construction. Then, among the activated nodes of the same layer in the original network, a node at the same level as the top node of the subnetwork and adjacent to it is prioritized. One of them is randomly selected as the trigger node A.

[0104] Use the trigger node A' and the top node of the FN area to create a third type of shadow node, and update the number of layers of the entire network node.

[0105] like Figure 8 As shown, a subnetwork structure is generated layer by layer using existing core nodes and those added to represent the FN region. Its vertex corresponds to node 25 in the figure. An activation node (node ​​21 in the figure) that is connected to and on the same layer as nodes 21-24 is selected as the trigger node. The trigger node is merged into node 25, and segment node 25 is changed to a third-type shadow node.

[0106] Based on the correlation coefficient-based neural network, the present invention utilizes the high degree of unity of its structure and features, and creates a unique function of full visualization of network features and structure and a direct interaction mode between the network and humans. It can directly increase, decrease and edit target features and network structure through human intervention, making its application scenarios more flexible.

Claims

1. A terrain monitoring method, characterized in that: The following steps are involved: Deploy targets at monitoring points and acquire image samples; Constructing a target detection model to detect the image sample, and outputting a first monitoring signal when the target detection model fails to detect the target or when the target detection model detects that the target deviates from an original position; The object detection model includes a plurality of nodes using correlation coefficients as activation functions, the nodes including a core node for extracting pixel feature units of an input image, and a multi-level segment node corresponding to pixel features of different scales of the input image, formed based on spatial relationships of the pixel feature units. The core nodes and the segment nodes form a tree structure that converges upward step by step from the core nodes, and outputs the object detection position through the segment node at the root. The object detection model is trained by the following method: Obtaining an initial image sample, and generating an object detection model based on the initial image sample; Acquire a training image sample and a valid pixel area of ​​the training image sample; Detecting the training image samples using the target detection model to obtain target detection positions; Based on the structure of the target detection model, backtracking to obtain the activated pixel area of ​​the target detection position; Calculating the pixel difference between the activated pixel area and the effective pixel area; Constructing a first type of shadow node for mutually replacing activation between nodes at the same level and at the same center position; a second type of shadow node for causing a set node that meets the activation condition to output an activated state when a set pixel area is activated; a third type of shadow node for causing a set node that meets the activation condition to output an inactivated state when a set pixel area is activated; and a third type of shadow node for causing a set node to be activated when the shadow node itself is activated. Dividing a plurality of false detection areas according to the pixel difference; Backtracking to obtain a first optimized node, where the first optimized node is the highest-level node on the branch where the first optimized node is located and is completely covered by the false detection area; The second-type shadow node is embedded between the first optimization node and its parent node to conditionally activate the pixel features corresponding to the first optimization node.

2. The method according to claim 1, characterized in that The embedding of the second-type shadow node between the first optimization node and its parent node to conditionally activate the pixel feature corresponding to the first optimization node includes: Determine whether there is a positive condition trigger node, where the positive condition trigger node is an inactivated node at the same level as the first optimization node and adjacent to the corresponding pixel feature; If the positive condition trigger node exists, construct a second-type shadow node for the positive condition, and use the first optimization node and the positive condition trigger node under the same parent node as child nodes of the second-type shadow node for the positive condition. When the second-type shadow node for the positive condition is anchored by the positive condition trigger node and the positive condition trigger node is activated, the second-type shadow node for the positive condition is activated. If the positive condition trigger node does not exist, a second type of shadow node with the negative condition is constructed to conditionally activate the pixel feature corresponding to the first optimization node.

3. The method according to claim 2, characterized in that The step of constructing the second type of shadow node of the negative condition includes: Acquire a number of activated nodes located at the junction of the false detection area and the valid pixel area as a shadow set; Constructing a false detection branch structure corresponding to pixel features of different scales in the false detection area, and using the top node of the false detection branch structure as a negative condition trigger node; A second type of shadow node of the negative condition is constructed with the shadow set and the false detection branch structure as child nodes.

4. The method according to claim 3, characterized in that It is set that when the nodes in the shadow set and the negative condition trigger node are both activated and anchored to each other, the second type of shadow node of the negative condition puts the negative condition trigger node into an inactivated state, so that the false detection branch structure is not detected by the target detection model.

5. The method according to claim 1, characterized in that After the step of calculating the pixel difference between the activated pixel area and the effective pixel area, the method further includes: Dividing a plurality of missed detection areas according to the pixel difference; Backtracking to obtain a second optimized node, where the second optimized node is the highest-level node on the branch where it is located and is completely covered by the missed-detection area; The second optimization node is replaced by the first type shadow node, the child nodes of the first type shadow node include the second optimization node and its replacement node, and when any one of the second optimization node and its replacement node is activated, the first type shadow node is activated.

6. The method according to claim 5, characterized in that The replacement node is one of the activated nodes at the same level as the second optimization node and having similar corresponding pixel features, and the center coordinates of the first shadow-like node, the replacement node and the second optimization node are all set to be the same.

7. The method according to claim 5, characterized in that After the step of dividing a plurality of missed detection areas according to the pixel difference, the method further includes: Determine the missing area of ​​the missed detection area in the target detection model; Constructing a missed detection branch structure corresponding to pixel features of different scales in the missing area; Selecting an activated node at the same level as the top node of the missed-detection branch structure as a joint triggering node; The associated trigger node and the top node of the missed-detection branch structure are used as child nodes of the third-type shadow node, and when the associated trigger node is activated and the child nodes of the third-type shadow node can all be anchored, the third-type shadow node is activated; The third type of shadow nodes are added to the target detection model.

Images