Underwater target detection method capable of filtering false target

Abstract

The invention belongs to the technical field of underwater target detection, and particularly relates to an underwater target detection method capable of filtering false targets. Comprising the following steps: acquiring an original sonar image based on a multi-beam sonar system; annotating the image; classifying and marking the underwater target; establishing a sonar image identification and classification model based on a YOLO network model; on the basis of underwater false target filtering of a multi-hypothesis data association technology, the underwater target detection method capable of filtering the false targets, which is provided by the invention, can realize dynamic optimization of an obstacle bounding box, rejection of the false targets and improvement of obstacle avoidance efficiency while ensuring the accuracy and the adaptive capacity of the detection method. Unnecessary and unreasonable obstacle avoidance actions are reduced, obstacle avoidance navigation parameters are optimized, the target false detection rate is reduced, and false targets are effectively filtered out.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater target detection, and particularly relates to an underwater target detection method capable of filtering out false targets. BACKGROUND

[0002] An underwater robot (including an underwater autonomous vehicle, an underwater tethered vehicle, a tracked underwater robot, etc.) needs to obtain path obstacle area information by using a forward-looking sonar system during underwater movement to complete obstacle avoidance, navigation and other functions. Through multi-channel acoustic data, coherent superposition and spatial filtering are performed to reconstruct an underwater scene image. In order to improve the visual recognition efficiency of human operators, the forward-looking sonar system usually performs a series of post-processing operations on the original acoustic data. These processing operations include noise filtering, contrast enhancement, pseudo-color mapping and other technical means. The finally output image presents a fan-shaped color visualization result, as shown in Figure 1 This processing method fully considers the perception characteristics of the human visual system, enhances the visual distinction of different intensity echo signals through color coding, and thus facilitates the rapid identification and judgment of underwater targets by operators. However, this processing strategy optimized for human vision has significant limitations in computer vision applications. First, although the filtering process can suppress noise, it may also cause loss of detail information, especially the feature information of weak targets may be excessively smoothed. Second, the pseudo-color mapping process is essentially a nonlinear transformation, which may destroy the continuity and monotonicity of the original acoustic intensity information, which adversely affects machine learning methods based on grayscale features. In addition, the fan-shaped image format has inherent defects in target labeling. The labeling box inevitably contains invalid pixels outside the fan-shaped image area. These redundant information will interfere with the feature learning process of the deep learning model, eventually leading to high false alarm rate of target recognition, generation of a large number of false targets, low obstacle avoidance success rate, generation of unreasonable obstacle avoidance path and other problems. SUMMARY

[0003] The purpose of the present application is to provide an optimized underwater target detection method capable of filtering out false targets. The method combines confidence-based dynamic parameter adaptive tracking to realize false target filtering in underwater target detection, and cooperates with a greedy matching strategy to reduce computational complexity and significantly improve the real-time performance of the system.

[0004] To achieve the above purpose, the present application adopts the following technical solutions.

[0005] A method for detecting underwater targets capable of filtering out false targets, comprising the following steps:

[0006] Step 1: Obtain an original sonar image based on a multi-beam sonar system;

[0007] Step 2: Image labeling;

[0008] The underwater target is classified and labeled, including: based on the geometric shape of the underwater target in the original sonar image, the classification and labeling is carried out, and the underwater target belonging to man-made facility structure or natural environment boundary is separately classified and labeled into two categories;

[0009] The optical image usually adopts a labeling system based on semantic categories, such as specific object categories like "vehicle" and "tree". In contrast, the forward-looking multi-beam sonar image has a unique "satellite view" feature. Due to the vertical opening angle limitation of the sonar, three-dimensional objects are compressed and projected on the imaging plane, losing height information. Based on this imaging feature, to improve the generalization ability and applicability of the recognition model, the present application proposes a labeling strategy based on geometric shapes. Common classifications include square objects, circular objects, and linear objects, as shown in Figure 2 .

[0010] In actual application scenarios, underwater robots may encounter various types of boundary structures during navigation. These boundaries can be divided into two categories: artificial boundaries, including offshore work platforms, port facilities, and other man-made structures; and natural environment boundaries, including underwater slopes, reefs, rock walls, and other terrain features. Given the importance of these boundary structures to the navigation safety of underwater robots, they should be labeled as independent target categories when building a target detection dataset to ensure that the detection model can accurately identify and avoid potential navigation risks, as shown in Figure 3 .

[0011] Step three, target detection; a sonar image recognition classification model is established based on a YOLO network model.

[0012] The present application learns sonar images based on the YOLO series network structure and generates a recognition model.

[0013] Step four, pseudo-target filtering; underwater pseudo-target filtering based on multi-hypothesis data association technology. To meet the actual needs of underwater autonomous navigation, the present application proposes an underwater pseudo-target filtering method based on multi-hypothesis data association technology. This design fully considers the special challenges and real-time requirements of underwater acoustic imaging.

[0014] The underwater acoustic imaging environment brings unique technical challenges to target tracking. First, the virtual scene phenomenon in acoustic imaging may trigger unnecessary obstacle avoidance response, so it is necessary to accurately identify and track the real target through a reliable target tracking method. Second, the complexity of the underwater environment leads to inherent defects in acoustic images: serious noise interference, low resolution, lack of target feature information, and blurred object contour. These characteristics make the target recognition accuracy of acoustic images generally lower than that of optical images, and multiple targets may appear in the same image, significantly increasing the tracking difficulty. In practical applications, these factors can easily lead to target tracking loss. To solve the above problems, the present application adopts a multi-hypothesis data association technology as the theoretical basis of the pseudo-target filtering method. The method calculates the spatial overlap and Euclidean distance between targets to establish a simple and efficient target matching mechanism. To ensure the stability and accuracy of pseudo-target filtering, the method comprehensively utilizes detection confidence and historical trajectory information to realize dynamic management of the motion cycle of various targets, while meeting the calculation resource constraints and real-time requirements of underwater robots. The method is efficient and can effectively avoid data processing bottlenecks.

[0015] Specifically includes:

[0016] D1, modeling of tracking target state

[0017] A seven-tuple representation of a tracking target full state mathematical model is established, where the state of the i-th tracking target at time t is ;

[0018] where, represents the bounding box scale of the i-th tracking target at time t, is the detection confidence of the i-th tracking target at time t, is the center point coordinate of the i-th tracking target at time t, is the class label of the i-th tracking target, is the unique identifier of the i-th tracking target, is the number of consecutive lost frames of the i-th tracking target at time t, is a history window with a length of W, used to store the historical state information of the tracking target, ;

[0019] D2, multi-feature fusion matching

[0020] D20, determining matching features,

[0021] Data association is the core problem of multi-target tracking, therefore the present application adopts a matching strategy based on multi-feature fusion, which comprehensively considers spatial overlap, motion consistency and detection confidence;

[0022] The overlap IoU refers to the overlap of the bounding boxes of the two i-th tracking targets and the jth detection target The degree of spatial overlap between the ith tracking target and the jth detection target is represented as ; is the intersection area; is the union area; and , ; , ;

[0023] is the center coordinate of the boundary box corresponding to the ith tracking target; and represent the length and height of the overlapping part of the detection box; is the center coordinate of the boundary box corresponding to the jth detection target, is the width of the detection box of the ith tracking target, is the height of the detection box of the ith tracking target, is the width of the detection box of the jth detection target, is the height of the detection box of the jth detection target;

[0024] Based on the IoU, the center point distance, and the detection confidence, the matching score between the state of the ith tracking target at time t-1 and the state of the jth detection target at time t is calculated ;

[0025] is the distance weight coefficient, is the confidence weight coefficient, is the Euclidean distance between the boundary boxes of the ith tracking target and the jth detection target, is the distance normalization function, is the detection confidence of the detection target at time t, represents the maximum allowed displacement distance of the ith tracking target;

[0026] D21, based on data association, adopts a greedy matching strategy to find the association scheme that maximizes the total matching score: ;

[0027] where A represents all possible matching association sets, is the optimal matching scheme, and the matching process is subject to the following constraints:

[0028] that is, the matching score needs to exceed a threshold value; is the matching threshold value;

[0029] (i) Each tracking target matches at most one detection;

[0030] (ii) Each detection is matched by at most one tracked target;

[0031] D22. To address different scenarios and target characteristics, the method designs a three-level adaptive parameter adjustment mechanism based on confidence level; specifically:

[0032] D221. Based on the target average confidence level Determine the confidence level and categorize it into high confidence, medium confidence, and low confidence levels based on the required confidence level.

[0033] For each confidence level L, define the corresponding parameter vector:

[0034]

[0035] The parameter vectors refer to: IoU threshold, etc. Maximum displacement distance Matching threshold Maximum number of lost frames Minimum number of tests and minimum detection rate ;

[0036] D222. Different parameter configurations and different matching score function weight coefficients are adopted according to the confidence levels of different objectives.

[0037] Low-confidence targets employ more stringent parameter configurations to reduce false positives, including using a larger IoU threshold. Larger matching threshold Fewer tests Smaller detection rate Smaller maximum displacement distance Smaller maximum number of lost frames One or more of them;

[0038] High-confidence targets employ more relaxed parameter configurations to improve tracking robustness, including using a larger maximum displacement distance. Larger maximum number of lost frames Smaller IoU threshold Smaller matching threshold Fewer tests and lower detection rate One or more of them;

[0039] The medium-confidence target uses parameter configurations intermediate between the low-confidence and high-confidence targets;

[0040] The weighting coefficients in the matching scoring function are dynamically adjusted based on the confidence level: for low-confidence targets... Take the larger value. Take the smaller value; for high-confidence targets, Take the smaller value. Take the larger value; for medium confidence targets, the weight coefficients of each item are the midpoint between the low confidence targets and the high confidence targets.

[0041] Weight coefficients in the matching scoring function , , The weighting is dynamically adjusted based on confidence levels, reflecting an important principle: "Higher confidence levels focus more on movement patterns; lower confidence levels focus more on morphological similarities." At low confidence levels, the precise location of the detected object is less trusted, with greater emphasis on whether it is "the same thing." This is because... This is reflected in the fact that at high confidence levels, there is greater trust in the accuracy of the detection, with a greater focus on whether the "movement is reasonable." This demonstrates that this scheme can make the most reasonable matching decision under different detection quality conditions.

[0042] D223. Employ a dual confirmation mechanism to reduce false detections and improve tracking reliability.

[0043] The first type is distance-based instant confirmation: when the identification model approaches the target to a relatively close range that would affect the navigation safety of the aircraft, it no longer performs tracking confirmation, but directly confirms the target;

[0044] The second method is statistical confirmation based on a sliding window; defined within a time window. Internal detection rate ;

[0045] A time window detection rate threshold is set. When the detection rate within the time window exceeds the threshold and the number of detections is greater than the minimum number of detections, the target is confirmed. This can be represented as:

[0046]

[0047] This refers to the total number of target detections;

[0048] Its beneficial effects are as follows:

[0049] The underwater target detection method proposed in this invention can filter out false targets. While ensuring the accuracy and adaptability of the detection method, it can achieve dynamic optimization of the obstacle bounding box, improve obstacle avoidance efficiency, reduce unnecessary and unreasonable obstacle avoidance actions, optimize obstacle avoidance navigation parameters, reduce the false target detection rate, and effectively filter out false targets. Attached Figure Description

[0050] Figure 1These are sonar images acquired by a forward-looking sonar system;

[0051] Figure 2 It is a classification and labeling based on the shape of the target in the sonar image;

[0052] Figure 3 It is a boundary classification and annotation based on the target boundary attributes in sonar images;

[0053] Figure 4 This is the online tracking result of the target detection method on the fishing net target (frame 50).

[0054] Figure 5 This is the online tracking result of the target detection method on the fishing net target (frame 51).

[0055] Figure 6 This is the online tracking result of the target detection method on the fishing net target (frame 52).

[0056] Figure 7 This is the online tracking result of the target detection method on the fishing net target (frame 53).

[0057] Figure 8 This is the online tracking result of the target detection method on the fishing net target (frame 54).

[0058] Figure 9 This is the online tracking result of the target detection method on the fishing net target (frame 55).

[0059] Figure 10 This is the online tracking result of the target detection method on the fishing net target (frame 56).

[0060] Figure 11 This is the online tracking result of the target detection method on the fishing net target (frame 57).

[0061] Figure 12 This is the curve showing the change in the false positive suppression rate of the target detection method for fishing net targets over time;

[0062] Figure 13 It is a curve showing the change in the detection counting time of the target detection method for fishing net targets over time;

[0063] Figure 14 Comparison of detection counts before and after target tracking during underwater navigation; Figure 15 This is a comparison of the regional suppression rates of target tracking methods during the underwater navigation phase;

[0064] Figure 16 This is a diagram illustrating the generation of a bounding box when confirmed obstacle information is received approximately 15 meters from the target.

[0065] Figure 17 This is a schematic diagram illustrating the gradual reduction of the directional bounding box radius based on a dynamic adjustment strategy;

[0066] Figure 18 This is a schematic diagram of obstacle avoidance based on dynamically adjusted back box;

[0067] Figure 19 These are schematic diagrams illustrating the tracking and obstacle avoidance of different types of continuous obstacles;

[0068] Figure 20 This is a view from the perspective of a drone tracking and avoiding obstacles such as fishing nets.

[0069] Figure 21 This is a view of a drone tracking and obstacle avoidance system targeting slopes with artificial boundary obstacles.

[0070] Figure 22 This is a view from a drone that tracks and avoids obstacles near natural boundary reefs.

[0071] Figure 23 This is a view of a drone tracking and avoiding obstacles on a regularly shaped metal cylinder. Detailed Implementation

[0072] To verify the effectiveness of the proposed underwater target detection method that can filter out false targets, an underwater target detection model based on YOLOv11 was first trained and evaluated based on the aforementioned basic scheme and steps. In terms of performance indicators, the model's precision and recall reached 98.5% and 98.0% respectively, the mAP50 index was stable at 99.1%, and the mAP50-95 reached 70.5%, indicating that the model has excellent detection capabilities for underwater targets.

[0073] Based on the trained detection model, this application systematically tests the effectiveness of the target tracking method.

[0074] Figures 4-11 This image demonstrates the actual effect of online tracking of the first target (fishing net) in frames 50-57. In the image, the target marked in pink (ID=0) is the real target confirmed by the tracking method, while the targets in red (ID=1 and ID=2) are unconfirmed targets, which are actually schools of fish that were misdetected as rectangular objects by the identification model (false targets).

[0075] The tracking method effectively eliminates these false detection targets through a virtual scene screening mechanism, ensuring that only confirmed real targets are transmitted to the navigation system to complete obstacle avoidance control.

[0076] To confirm the false detection suppression performance of the method in this application, a detailed quantitative analysis is performed below.

[0077] Figures 12-13This demonstrates the performance of the method proposed in this application during single target tracking. Figure 12 The temporal evolution of the false positive suppression rate was revealed: during the initialization phase (0-20 frames), the system exhibited significant fluctuations due to parameter adjustments; subsequently, it converged rapidly and remained stable, with the average suppression rate maintained at a low level of 4.5%, showing only slight perturbations in the 50-70 frame range, demonstrating the stability of the method. Figure 13 The comparison of the number of detections shows that although there are two original detections in the initial frame, the method always keeps the number of confirmed detections to one, and from the 10th frame onwards, it achieves complete consistency between the original detections and the confirmed detections, proving the accuracy of the method.

[0078] Figures 14-15 The method's overall performance throughout the entire underwater navigation process was further demonstrated, from the statistical analysis of the number of detections ( Figure 14 As can be seen, the total number of tests decreased from 1462 to 1087, with an overall suppression rate of 25.6%. Among them, the suppression effect of low-confidence tests was particularly significant, decreasing sharply from 667 to 337, a reduction of nearly 50%; while the changes in medium- and high-confidence tests were relatively gradual. Figure 15 The partitioned suppression rate analysis revealed the method's intelligent screening strategy: the suppression rate in the low-confidence interval was as high as 49.5%, far exceeding the reference threshold of 30%; while the suppression rates in the medium-confidence (6.5%) and high-confidence (2.8%) intervals were significantly lower than the threshold.

[0079] As shown above, based on the differentiated processing strategy of this application, while actively filtering out low-quality detections, high-confidence detections are carefully retained to prevent missed detections. In summary, this target tracking method successfully achieves a balance between false detection suppression and detection integrity, providing reliable target information for autonomous obstacle avoidance.

[0080] Once the obstacle management system receives the target information confirmed by the tracking method, it initiates the obstacle avoidance procedure.

[0081] Figures 16-18 It shows in detail the entire process of avoiding the first target (the fishing net).

[0082] like Figure 16 As shown, when confirmed obstacle information is received at a distance of approximately 15 meters from the target, this method can immediately generate a directional bounding box with a radius of approximately 11 meters around the obstacle. This bounding box not only defines the safety boundary but also provides explicit geometric constraints for path planning. Based on obstacle avoidance rules, a detour path is planned along the edge of the bounding box.

[0083] However, the complexity of the actual underwater environment poses a significant challenge to its maneuverability. In initial obstacle avoidance attempts, due to water currents and its own maneuverability limitations, it was unable to complete the detour along the preset path. To address this issue, this method demonstrates its adaptive characteristics. Figure 17 As shown, after detecting a maneuvering difficulty, the system gradually reduces the radius of the directional bounding box based on the dynamic adjustment strategy of this application, creating a more flexible obstacle avoidance space. This gradual adjustment continues until the optimal balance between current maneuverability and obstacle avoidance requirements is found.

[0084] Ultimately, as Figure 18 As shown, the obstacle avoidance maneuver was successfully completed through dynamic optimization of the bounding box. Throughout the process, the obstacle avoidance rules not only ensured the safety of obstacle avoidance but also fully considered the actual maneuver constraints, achieving an organic combination of theoretical methods and engineering practice. The successful application of this adaptive obstacle avoidance strategy significantly improves the autonomous navigation capability in complex underwater environments.

[0085] Figure 19 The entire trajectory of the navigation process, which involved successively avoiding four different types of obstacles, was recorded. Figures 20-23 This is an aerial view taken by a drone.

[0086] As can be clearly observed from the navigation trajectory, the vehicle successfully identified and avoided four confirmed obstacles (marked with blue dots) while navigating along the preset path (blue dashed line). The first obstacle was a fishing net located at the beginning of the route, which the vehicle avoided by maneuvering to the right. This process fully verified the effectiveness of the aforementioned adaptive bounding box strategy.

[0087] Subsequently, the vehicle encountered a second obstacle—a slope. Faced with this large-scale terrain obstacle, the vehicle performed a more significant maneuver, with its trajectory showing a clear arc-shaped deviation, demonstrating the improved design's adaptability to obstacles of different scales.

[0088] After successfully avoiding the slope, the vehicle detected a third obstacle—a reef—midway through its flight path. Because reefs typically have irregular shapes and potential extensions, the vehicle adopted a relatively conservative obstacle avoidance strategy, executing a full "S"-shaped maneuver to maintain a safe distance from the reef. This cautious obstacle avoidance behavior reflects the system's intelligent assessment of the risk level of different obstacle types.

[0089] Finally, before approaching the target point, the aircraft detected the fourth obstacle—a metal cylinder. As a small, regularly shaped artificial target, the system accurately assessed its boundaries and executed precise, minor avoidance maneuvers, maximizing heading efficiency while ensuring safety.

[0090] Throughout the entire flight, although the actual trajectory of the aircraft (solid red line) deviated from the preset path multiple times due to obstacle avoidance, it always maintained good heading tracking ability and was able to quickly return to the predetermined route after completing obstacle avoidance.

[0091] The above continuous obstacle avoidance test verified the practical application effect of the technical solution of this application in target recognition, tracking confirmation, path planning and other functions, and proved the robustness and practicality of the solution of this application in the face of diverse underwater obstacles.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for detecting underwater targets that can filter out false targets, characterized in that, Includes the following steps: Step 1: Obtain raw sonar images based on the multibeam sonar system; Step 2: Image annotation; The underwater targets are classified and labeled, including: classifying and labeling the underwater targets based on the geometry of the underwater targets presented in the original sonar images, and classifying and labeling underwater targets belonging to man-made structures or natural environmental boundaries into two separate categories; Step 3: Target detection; Establish a sonar image recognition and classification model based on the YOLO network model; Step 4: False Target Filtering. Underwater false target filtering based on multi-hypothesis data association technology specifically includes: D1. Target state modeling Establish a mathematical model of the full state of the tracking target represented by a 7-tuple, where the state of the i-th tracking target at time t is... ; in, This represents the bounding box scale of the i-th tracked target at time t. Let be the detection confidence score of the i-th tracked target at time t. Let be the coordinates of the center point of the i-th tracked target at time t. Let i be the category label for the i-th tracked target. A unique identifier for the i-th tracked target. Let be the number of consecutive frames lost at time t for the i-th tracked target. A history window of length W is used to store the historical state information of the tracked target. ; D2, Multi-feature fusion matching D20. Determine matching features, including spatial overlap, motion consistency, and detection confidence. IoU (Intersection over Union) refers to the bounding box of the i-th tracked target. and the j-th detection target The degree of spatial overlap between them is expressed as ; It refers to the area of ​​intersection; It refers to the area of ​​the union; and , ; , ; This refers to the center coordinates of the bounding box corresponding to the i-th tracked target; and This represents the length and height of the overlapping portion of the detection boxes; This refers to the center coordinates of the bounding box corresponding to the j-th detected target. This refers to the width of the detection box for the i-th tracked target. This refers to the height of the detection box of the i-th tracked target. This refers to the width of the detection box for the j-th target. This refers to the height of the bounding box of the j-th detected target. Based on IoU, center point distance, and detection confidence, calculate the state of the i-th tracked target at time t-1. The state of the target at time t is tracked by the j-th detected target. Match score between ; This is the distance weighting coefficient. The confidence weighting coefficient is... Let be the Euclidean distance between the bounding boxes of the i-th tracked target and the j-th detected target. This is the distance normalization function. It refers to the detection confidence of the target at time jt. This represents the maximum allowable displacement distance of the i-th tracked target; D21. Based on data association, employ a greedy matching strategy to find the association scheme that maximizes the total matching score: ; Where A represents the set of all possible matching associations, For the optimal matching scheme, the matching process is subject to the following constraints: That is, the matching score must exceed the threshold; This refers to the matching threshold; (i) Each tracked target is matched with at most one detection; (ii) Each detection is matched by at most one tracked target; D22. Set up a three-level adaptive parameter adjustment mechanism based on confidence level, specifically referring to: D221. Based on the target average confidence level Determine the confidence level and categorize it into high confidence, medium confidence, and low confidence levels based on the required confidence level. For each confidence level L, define the corresponding parameter vector: The parameter vectors refer to: IoU threshold, etc. Maximum displacement distance Matching threshold Maximum number of lost frames Minimum number of tests and minimum detection rate ; D222. Different parameter configurations and different matching score function weight coefficients are adopted according to the confidence levels of different objectives. Low-confidence targets employ more stringent parameter configurations to reduce false positives, including using a larger IoU threshold. Larger matching threshold Fewer tests Smaller detection rate Smaller maximum displacement distance Smaller maximum number of lost frames One or more of them; High-confidence targets employ more relaxed parameter configurations to improve tracking robustness, including using a larger maximum displacement distance. Larger maximum number of lost frames Smaller IoU threshold Smaller matching threshold Fewer tests and lower detection rate One or more of them; The medium-confidence target uses parameter configurations intermediate between the low-confidence and high-confidence targets; The weighting coefficients in the matching scoring function are dynamically adjusted based on the confidence level: for low-confidence targets... Take the larger value. Take the smaller value; for high-confidence targets, Take the smaller value. Take the larger value; for medium confidence targets, the weight coefficients of each item are the midpoint between the low confidence targets and the high confidence targets.

2. The underwater target detection method capable of filtering out false targets according to claim 1, characterized in that, Step D22 further includes: D223. Employ a dual confirmation mechanism to reduce false detections and improve tracking reliability. The first type is distance-based instant confirmation: when the identification model approaches the target to a relatively close range that would affect the navigation safety of the aircraft, it no longer performs tracking confirmation, but directly confirms the target; The second method is statistical confirmation based on a sliding window; defined within a time window. Internal detection rate ; A time window detection rate threshold is set. When the detection rate within the time window exceeds the threshold and the number of detections is greater than the minimum number of detections, the target is confirmed. This can be represented as: This refers to the total number of target detections.

3. The underwater target detection method capable of filtering out false targets according to claim 1, characterized in that, In step two, the classification based on geometric shape includes at least: square objects, circular objects, and linear objects.

Images