Clustering noise reduction device, method, chip, event imaging device and electronic equipment

Abstract

The present invention discloses a clustering noise reduction device, method, chip, event imaging device and electronic device. In order to solve the technical problem of the existence of noise in the event imaging process, the noise reduction scheme of the present invention includes receiving new events in real time, and continuously storing the new events in the address corresponding to the storage space, the storage space changes dynamically and is updated in real time; determining the clustering range in the storage space, determining whether the new event forms an effective clustering relationship with other events in the storage space, and judging whether the event belongs to a noise event or a valid event according to the clustering relationship. Through these technical means, the noise reduction scheme disclosed by the present invention can not only filter out ordinary noise, but also filter out special noise. While filtering out noise, it can retain valid events to the greatest extent, with little damage to valid events, and has technical effects such as fast response speed, low delay, and low power consumption.

Description

Technical Field

[0001] The present invention relates to a clustering noise reduction device, method, chip and electronic equipment, and in particular to a device, method, chip and electronic equipment for performing clustering noise filtering on pulse events output by an event imaging device. Background Art

[0002] Event imaging does not capture images at a fixed rate, but asynchronously measures the brightness changes of each pixel and outputs a series of events. Compared with traditional cameras, event cameras offer attractive features: low power consumption, strong real-time performance, large dynamic range, high resolution, etc. Common event imaging devices include dynamic vision sensors DVS, which are visual sensors suitable for neuromorphic and brain-like computing. However, due to the limitations of current technology, the actual output of DVS has various types of noise. Researchers from various countries have also done a lot of research in this frontier field, but most filtering technologies have problems such as long processing cycle, low noise reduction efficiency, and low effective event retention rate.

[0003] One of the core indicators of the noise reduction method for event imaging devices is the noise reduction speed. This is because when the target in the field of view shifts, the event data arrives quickly and in large quantities. If it is not a high-speed noise reduction algorithm, it lacks practical application value.

[0004] A common noise reduction scheme is correlation filtering, which determines whether other events occur within a period of time or range when the current event occurs. If so, it is considered a valid event. In some schemes, the timestamps of eight neighboring pixels around each pixel are stored, and the new timestamp overwrites the previous timestamp. Then, it is determined whether the difference between the current timestamp and the previous timestamp is greater than a threshold. If it is greater than the threshold, it is considered a noise event and filtered out. However, in this type of method, the filtering effect is not ideal, and only low-frequency background noise can be filtered out. When there is background noise near the hot pixel noise, the noise reduction effect will be affected. At the same time, this type of method does not fully utilize the spatial range and will lose spatial information.

[0005] In order to filter out hot pixel noise, a noise reduction scheme is to perform two-step filtering, such as prior art 1: CN111770290A (Changchun Institute of Optics, Fine Mechanics and Physics, Lv Hengyi, etc.), firstly, coarse filtering is performed, and a loose threshold is used to filter out background noise, and then fine filtering is performed to determine whether there are enough related events around the first event. Specifically, during coarse filtering, the number of events within a 5×5 range near the first event is first calculated, and events with a number less than the threshold are considered to be noise events and filtered out. Then fine filtering is performed, and a stricter 3×3 range is used to determine whether there are other related events near the first event. If there are, the event is retained, otherwise, the event is filtered out. However, there are two important indicators for measuring the quality of noise reduction effect, namely, the number of events that are originally noise, how many are missed by filtering, and how many events that are not noise are mistakenly deleted. Both steps of this method have a high dependence on the value of the spatial range. If the spatial range is too large, noise events will be missed, and the filtering effect will be poor; if the spatial range is too small, valid events will be misjudged, and valid events will be filtered out, affecting the imaging quality. At the same time, this method is highly complex and time-consuming, and does not process special noise (such as black block noise), so its practical value is relatively insufficient.

[0006] In view of the shortcomings of the above-mentioned event imaging denoising technology, the present invention proposes a clustering denoising technology, which has high real-time performance, low energy consumption, simplicity and effectiveness, and is easy to implement in hardware, with strong flexibility and wide applicability without affecting the excellent performance of existing event imaging. Summary of the invention

[0007] In order to solve or alleviate some or all of the above problems, the present invention proposes a clustering noise reduction scheme and performs noise filtering according to the characteristics of various noise events. Specifically, it is achieved in the following ways:

[0008] A clustering noise reduction device receives a plurality of events output by an event imaging device, wherein data information of the plurality of events includes at least coordinate information and time information, and the plurality of events includes a first event. The clustering noise reduction device includes: a first storage space configured to at least store time information of events; a mapping module configured to store at least time information of the first event in a first storage unit located in the first storage space and corresponding to the coordinate information of the first event according to the coordinate information of the first event; a clustering operation module configured to determine a clustering range of the first event in the first storage space according to the coordinate information of the first event; and determine whether the first event forms a valid clustering relationship with events in the clustering range according to the relationship between the time information of the events stored in the clustering range and the time information of the first event, and if so, determine that the first event is a valid event, and otherwise determine that the first event is a noise event.

[0009] In a certain class of embodiments, after at least time information of some events among the several events generated later than the first event is stored in the first storage space, whether the first event forms a valid clustering relationship with the events within the clustering range is determined based on the relationship between the time information of the events stored within the clustering range and the time information of the first event; or, before or after at least time information of the first event is stored in the first storage unit located in the first storage space and corresponding to the coordinate information of the first event, immediately based on the relationship between the time information of the events stored within the clustering range and the time information of the first event, whether the first event forms a valid clustering relationship with the events within the clustering range is determined.

[0010] In a certain class of embodiments, the clustering range of the first event is: in the event imaging device, an address range in the first storage space mapped to event imaging units that are physically adjacent to the event imaging unit that generates the first event.

[0011] In a certain class of embodiments, the clustering operation module includes a first threshold judgment module and a second threshold judgment module; the first threshold judgment module is configured to: based on the time information of the event, respectively calculate the time interval between the first event and each event within the cluster range, compare the calculated time interval with the first threshold, and obtain the time relationship value between the first event and each event within the cluster range; add the time relationship value between the first event and each event within the cluster range to obtain the sum of the cluster time relationship values; the second threshold judgment module is configured to: compare the sum of the cluster time relationship values ​​with the second threshold, and determine whether the event is a noise event or a valid event based on the comparison result.

[0012] In a certain class of embodiments, the clustering operation module includes a third threshold judgment module; the third threshold judgment module is set with a third threshold, wherein the third threshold is a time threshold or a frequency threshold; if the third threshold is a time threshold, if the average time interval or / and the maximum time interval of all events within the clustering range is less than the third threshold, then the event is determined to be a noise event; if the third threshold is a frequency threshold, if the average frequency or / and the minimum frequency of all events within the clustering range is greater than the third threshold, then the event is determined to be a noise event.

[0013] In certain embodiments, the first threshold determination module: calculates the time relationship value between the first event and each event within the cluster in parallel; or / and, adds the time relationship value between the first event and each event within the cluster in parallel.

[0014] In a certain type of embodiment, if the time interval is larger than the first threshold, the time relationship value is 0; if the time interval is smaller than the first threshold, the time relationship value is 1; if the sum of the cluster time relationship values ​​is larger than the second threshold, it is determined to be a valid event, otherwise it is a noise event.

[0015] A clustering noise reduction method, the clustering noise reduction method comprising the following steps: receiving a number of events output by an event imaging device, wherein the data information of the number of events at least includes coordinate information and time information, and the number of events includes a first event; based on the coordinate information of the first event, storing at least the time information of the first event in a first storage unit located in a first storage space and corresponding to the coordinate information of the first event; based on the coordinate information of the first event, determining the clustering range of the first event in the first storage space; based on the relationship between the time information of the events stored in the clustering range and the time information of the first event, determining whether the first event forms a valid clustering relationship with the events in the clustering range, and if so, determining the first event to be a valid event, otherwise determining the first event to be a noise event.

[0016] In a certain class of embodiments, after at least time information of some events among the several events generated later than the first event is stored in the first storage space, whether the first event forms a valid clustering relationship with the events within the clustering range is determined based on the relationship between the time information of the events stored within the clustering range and the time information of the first event; or, before or after at least time information of the first event is stored in the first storage unit located in the first storage space and corresponding to the coordinate information of the first event, immediately based on the relationship between the time information of the events stored within the clustering range and the time information of the first event, whether the first event forms a valid clustering relationship with the events within the clustering range is determined.

[0017] In a certain class of embodiments, the clustering range of the first event is: in the event imaging device, an address range in the first storage space mapped to event imaging units that are physically adjacent to the event imaging unit that generates the first event.

[0018] In a certain class of embodiments, based on the time information of the event, the time interval between the first event and each event within the cluster range is calculated respectively, and the calculated time interval is compared with a first threshold to obtain the time relationship value between the first event and each event within the cluster range; the time relationship values ​​between the first event and each event within the cluster range are added to obtain the sum of the cluster time relationship values; the sum of the cluster time relationship values ​​is compared with a second threshold, and the event is determined to be a noise event or a valid event based on the comparison result.

[0019] In a certain class of embodiments, a third threshold is set, wherein the third threshold is a time threshold or a frequency threshold; if the third threshold is a time threshold, if the average time interval and / or the maximum time interval of all events within the cluster range is less than the third threshold, the event is determined to be a noise event; if the third threshold is a frequency threshold, if the average frequency and / or the minimum frequency of all events within the cluster range is greater than the third threshold, the event is determined to be a noise event.

[0020] In certain embodiments, the time relationship value between the first event and each event within the cluster is calculated in parallel; or / and, the time relationship value between the first event and each event within the cluster is added in parallel.

[0021] In a certain type of embodiment, if the time interval is larger than the first threshold, the time relationship value is 0; if the time interval is smaller than the first threshold, the time relationship value is 1; if the sum of the cluster time relationship values ​​is larger than the second threshold, it is determined to be a valid event, otherwise it is a noise event.

[0022] A clustering denoising method is disclosed. The method receives a first event output by an event imaging device, obtains data information of the first event, and the data information at least includes coordinate information; wherein the coordinate information is the position of the first event when a detection plane of the event imaging device is generated; according to the coordinate information, the first event is mapped to a corresponding first storage unit in a first storage space; the clustering range of the first event is determined in the first storage space; and through a clustering operation, the events within the clustering range are classified to obtain at least a valid event set and a noise event set.

[0023] In certain embodiments, the clustering operation adopts K-means clustering and / or mean shift clustering.

[0024] A chip comprises the clustering noise reduction device as described in any one of the preceding items, or the chip applies the clustering noise reduction method as described in any one of the preceding items.

[0025] In certain embodiments, the clustering noise reduction device is implemented in a dedicated interface circuit; or is implemented as an FPGA circuit; or is implemented by a combination of software and hardware; or the clustering noise reduction method is implemented in a host computer through software.

[0026] An event imaging device, comprising the clustering denoising device as described in any one of the preceding items, or the event imaging device applies the clustering denoising method as described in any one of the preceding items.

[0027] An electronic device comprises an event imaging device and a clustering noise reduction device as described in any one of the preceding items, wherein the clustering noise reduction device is configured to perform noise filtering on a plurality of events output by the event imaging device.

[0028] The noise event determined by the second threshold determination module is an event generated by random noise or / and hot pixel noise. The noise event determined by the third threshold determination module is an event generated by black noise or / and black block noise.

[0029] The present invention combines the structural principle of the event imaging device and proposes a real-time, clustering noise reduction technology to address the difficulties faced by current noise reduction technology. It can achieve high-performance noise filtering in a shorter processing cycle, thereby effectively improving the performance of DVS effective event identification, DVS image algorithm processing, and DVS detection structure power consumption.

[0030] Compared with the prior art, some or all of the embodiments of the present invention have the following beneficial effects:

[0031] (1) Based on the characteristics of event imaging, the present invention performs real-time noise reduction processing with fast response speed, low delay and strong real-time performance.

[0032] (2) The present invention does not require frame compression, completely retains the time attributes of events, directly filters noise from the three-dimensional pulse event stream, avoids the loss of valid events, supports parallel processing, and is easy to implement in hardware.

[0033] (3) The present invention clusters events based on the temporal and spatial distribution relationship between each event and the event that occurred before the DVS detection plane. The noise reduction method is simple and effective. It can filter out noise while ensuring accuracy and retain valid events. It has important reference and guiding significance for improving the performance of DVS cameras.

[0034] (4) The processing of the present invention can not only filter out common noise (background random noise, hot pixel noise), but also filter out special noise (black block noise, black noise), with good filtering performance, strong flexibility and wide applicability.

[0035] (5) The present invention can maximize the retention of valid events while filtering, with little damage to valid events and a low false alarm probability.

[0036] (6) The present invention effectively filters out unwanted environmental noise, and when there is no moving object within the DVS field of view, the first event is not generated.

[0037] (7) The present invention can effectively filter out various noise events. When DVS is activated, the motion trajectory of the target object is clear and intact, and the thickness of the entire motion trajectory is basically consistent. In the process of noise reduction, no valid events will be accidentally damaged, and the noise reduction effect is good.

[0038] More beneficial technical effects will be described in the subsequent preferred embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of a scene of the clustering denoising method of the present invention;

[0040] Figure 2 It is a schematic diagram of storage space mapping according to a preferred embodiment of the present invention;

[0041] Figure 3 It is a schematic diagram of storage space corresponding to the DVS detection plane in a preferred embodiment of the present invention;

[0042] Figure 4 It is a schematic diagram of cluster determination in a preferred embodiment of the present invention;

[0043] Figure 5 The storage space corresponding to the first event and the clustering range in the embodiment of the present invention;

[0044] Figure 6 A schematic diagram of the distribution of events having a time relationship with the first event within a clustering range according to an embodiment of the present invention;

[0045] Figure 7 It is a schematic diagram of determining special noise in a preferred embodiment of the present invention;

[0046] Figure 8 is a flow chart of clustering noise reduction processing provided by an embodiment of the present invention;

[0047] Fig. 9 is a simplified diagram of the clustering denoising process provided by an embodiment of the present invention;

[0048] Fig.10a It is the black noise spectrum existing in DVS when there is no valid event before noise reduction;

[0049] Fig.10b It is a black noise spectrum existing in DVS after noise reduction using a preferred embodiment of the present invention;

[0050] Fig.11a It is the black block noise spectrum existing in DVS when there is no valid event before noise reduction;

[0051] Fig.11b It is a black block noise spectrum existing in DVS after noise reduction using a preferred embodiment of the present invention;

[0052] Fig.12a It is the event spectrum when DVS is activated before filtering;

[0053] Figure 12b is an event graph of another embodiment when the DVS is activated after noise reduction;

[0054] Fig.13 It is a schematic diagram of back-end processing of a pulse neural network processor provided in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0055] Other technical solutions and details not disclosed in detail below are generally conventional technical means / common knowledge in the field, and due to space limitations, the present invention will not introduce them in detail.

[0056] Unless it means division, " / " in any position in the present invention means logical "or". The serial numbers such as "first", "second", "third" and so on in any position in the present invention are only used as distinguishing marks for description, and do not imply an absolute order in time or space, nor do they imply that the terms prefixed with such serial numbers and the same terms prefixed with other attributives must necessarily refer to different things.

[0057] When describing the existence or inclusion of a certain step or module at any position in the present invention, it does not imply that such existence is exclusive and unique. Those skilled in the art can obtain other technical solutions based on the technical solutions disclosed in the present invention with the assistance of other technical means. Such solutions do not deviate from the concept of the present invention and are within the protection scope of the present invention.

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0059] Terminology explanation:

[0060] Event imaging: sensing changes in ambient light or capturing changing scenes within the field of vision, and generating a series of events, called an event stream or event packet.

[0061] Random noise: This noise is randomly generated in the N×N two-dimensional plane of the event imaging device, so that its continuous events are evenly distributed in space, and usually there are no consecutive points appearing in the same small plane area.

[0062] Hot Pixel Noise: refers to the fact that some pixel mechanisms of event imaging devices have certain problems, which will cause frequent false triggers and form a fixed noise source. The hot spot noise triggering time interval is usually at the millisecond level, and there will be other noises or events in the interval, so there will not be several continuous noises appearing in the same small plane area.

[0063] Black block noise: The structure of some event imaging devices is too sensitive to objects that are all or partly black in the detection range, which will cause false triggers in the entire area and generate frequent events. This situation is called black block noise in this article. Its characteristic is that the trigger frequency is much higher than the trigger frequency of normal events, and the trigger period is usually less than 10us.

[0064] Black noise: refers to the fact that some event imaging device structures are very sensitive in no light or low light conditions, and will continuously trigger falsely. Its noise covers the entire event imaging device detection range, causing the event imaging device detection structure to work at full load, greatly affecting the stability of the system. Its characteristic is that events are continuously output without intervals, that is, there is almost no time interval between two events.

[0065] Clustering: gather data within a certain range and divide the data set into multiple subsets based on certain classification rules, where the data in each subset has a high degree of intrinsic similarity. The present invention gathers events within a certain range and classifies them to obtain a noise event set and a valid event set.

[0066] Cluster range: also referred to as cluster space or spatial cluster in this article. A set of event imaging units that are physically adjacent to a certain event imaging unit that generates the first event in the event imaging device, or / and the address range / set after it is mapped in the storage space W. Corresponding to the event imaging device, the adjacent range can be square, rectangular, prototype, etc. For example, Figure 2 As shown, the clustering range is the gray area in the storage space W. The clustering range may not include the storage unit corresponding to the first event (because the first event and itself must belong to the same category in time and space, but it does not violate the principle to also classify the storage unit into the clustering space). Figure 6 As shown in , the clustering range is a 17×17 pixel point square matrix around the storage unit corresponding to the first event as the center.

[0067] Temporal relationship value: a measurement value that characterizes the distance between two moments. If it is less than a certain threshold, it is considered that the two moments are close in time, for example, it can be represented by a value of 1; otherwise, it is considered that the relationship is far, for example, it can be represented by a value of 0. The sum of cluster temporal relationship values ​​is the sum of several temporal relationship values ​​within the cluster range.

[0068] Spiking neural network (SNN): The third generation of neural network, the information transmission inside the neuron is completed by a pulse sequence (or event stream), the pulse sequence is a time series composed of discrete pulse time points, and the data set is composed of spatiotemporal event streams, which has the advantage of ultra-low energy consumption in hardware circuits. SNN neurons are more bionic than ANN neurons, with rich time information and neural dynamics characteristics (threshold emission characteristics), and have achieved excellent performance in efficiently processing complex, sparse and noisy spatiotemporal information.

[0069] Neuron: At the chip level, it refers to hardware (such as circuits) or software or / and a combination of hardware that simulates the working principles of biological neurons.

[0070] Pulse neural network processor: Different from the traditional von Neumann architecture, it does not have the traditional concept of data and instructions. It is a new computing paradigm. It is equipped with a pulse neural network, which transmits information by using the release time / frequency of the pulse sequence, and has a real simulation process such as membrane voltage accumulation. It is also called a neuromorphic chip, a neuromorphic chip or a brain-like chip.

[0071] This paper will classify the noise into four types according to the noise characteristics of event imaging: random noise, Hot Pixel noise, black block noise and black noise. Based on the characteristics of each type of noise, this paper clusters and determines the events according to the time and spatial distribution characteristics of each event within a limited time and space, so as to retain the real and valid events and filter out the above-mentioned noise problems.

[0072] Event imaging does not generate frame images. Each pixel independently and asynchronously measures the change in its pixel brightness. As long as there is a change in the brightness of a pixel in the field of view, an event is generated. There is no need to wait, and it has strong real-time performance. Since all events occur independently and asynchronously, even the smallest time interval cannot occur completely at the same time. Therefore, the moments of events generated by the event imaging device (which can be represented by timestamps) are all different. Event imaging devices include but are not limited to DVS (dynamic vision sensor) and PET (positron emission tomography) readout devices. Taking DVS as an example below, when collecting events, only one changing pulse event is output at each moment (which can be represented by a timestamp). Therefore, the detection plane of DVS changes dynamically.

[0073] Since event imaging devices do not capture static images or generate grayscale frames, the event stream generated is a completely new form of data, so existing noise reduction technologies are difficult to apply. Existing methods compress event streams, such as compressing events within a time interval to generate event frames. The format of the frames can be accepted by traditional artificial neural networks, giving full play to the advantages of existing noise reduction algorithms. However, this method violates the essence of event imaging and introduces the concept of frames. The compression dimension means information loss, and real-time performance is also affected. Reduced real-time performance means delays, motion blur and other problems.

[0074] When the brightness value of a certain pixel changes during event imaging, several events are generated in real time, including the first event. The noise reduction scheme of the present invention fully adapts to the characteristics of event imaging, performs noise reduction processing on the first event generated in real time, and has an extremely fast response speed. The first event in the present invention can be the latest and last event generated among the several events, or it can be an event before the latest and last event. For the former, as soon as the event is generated, it is immediately determined whether it is a noise event or a valid event based on the clustering judgment result. For the latter, after the event is generated, it is not immediately determined whether it is a noise / valid event, but the event generated after being updated is allowed to store its time information in the first storage space, and then the clustering judgment is completed, so as to avoid the newly generated event being mistakenly judged as a noise event.

[0075] Based on the above noise characteristics of DVS, the present invention proposes a clustering noise reduction method for the distribution characteristics of DVS events, which directly filters noise from the pulse event stream without frame compression, thus avoiding the loss of valid events containing continuous motion information. Figure 1 As shown in the figure, each square in the plane stores the latest sampled event of the corresponding pixel point. The black square is the storage coordinate corresponding to the new input event, and the gray square is the event in space that has a certain event association with the new input event. The specific process is as follows:

[0076] Step S1: DVS outputs a first event in real time. Specifically, a plurality of events output by an event imaging device are received, wherein data information of the plurality of events at least includes coordinate information and time information, and the plurality of events includes a first event.

[0077] The data information of the DVS output event includes information such as coordinates and time. The data information of the event is also referred to as event information or event data in this article. Time information includes the time when the event occurs, which can be specifically represented by a timestamp. The data of the DVS output event may also include polarity, which is used to indicate the form of the change in light intensity at the pixel point where the event occurs, whether it is brighter or darker. As an example, the event data of a pixel event can be represented based on Address-Event Representation (AER), or it can be represented in other ways including coordinate information and time information. In addition, for the expression of events, the more dimensions, the more resources occupied. In some cases, the polarity of the event can be omitted.

[0078] The coordinate information includes the location where the event occurs in the N×K detection plane. Set the DVS detection plane coordinates to the XY coordinate system. For example, the location where the event occurs is (i, j), where N and K are positive integers, representing the coordinate range of the DVS, N×K can represent the resolution of the DVS, and i and j are also positive integers. The detection plane of the DVS can be a square or a rectangle, or other shapes, such as a circle, etc. The representation of the position information can also be other representations, such as polar coordinates, etc., which are not limited in the present invention.

[0079] Step S2: Map the first event output by the DVS to a corresponding storage space. Specifically, according to the coordinate information of the first event, store at least the time information of the first event in a first storage unit located in the first storage space and corresponding to the coordinate information of the first event.

[0080] Storage space or memory is used to dynamically store the event stream output by the DVS. For example, Figure 2 As shown in , there is an N×K×M storage space W corresponding to the DVS N×K detection plane, where N is the X coordinate range of the DVS, K is the Y coordinate range of the DVS, and M corresponds to the size of the storage area required to store the data information of a DVS event, that is, the data information of each event is M bits, or the timestamp bit width. N, K and M can be adjusted as needed and are all positive integers. Storing event information in the storage space is sometimes also called storing events in the storage space. The storage space can be a spatially ordered register, a RAM storage unit, a disk storage space, etc.

[0081] Corresponding to the event location information, each point in the storage space W can be represented in a corresponding manner. When the DVS has the first event E output, according to the coordinates or position (i, j), the first event E is mapped to the corresponding storage position W (i, j) in the W space. Since the storage address of the W space has already represented the coordinate or position information, in one embodiment, only the data information containing the time and / or polarity characteristics of the first event E can be stored in the storage position W (i, j). This method further reduces the amount of data, is faster, more concise and effective, and consumes less resources. At the same time, while storing data information with characteristics such as time and / or polarity, those skilled in the art can decide whether to store the location coordinate information according to actual needs. Of course, the increase in the amount of data means an increase in resource overhead and an increase in computational complexity, which will affect real-time performance.

[0082] As the events generated by DVS are continuously received, the event stream will be completely distributed in the W space, so that the storage location W(i, j) corresponding to the first event E will inevitably retain the event information that occurred before the first event arrived. The data of the first event E can be used to overwrite the previous event data, that is, each time a new first event E is stored, the original event data at the address is overwritten, thereby ensuring that the events stored in each storage location W(i, j) are the latest events that occurred or were collected in the area.

[0083] Therefore, for the W space, the memory stores the event stream output by the DVS in order, and each storage location (unit) stores the latest event collected by the pixel point. In order to achieve the goal of real-time and accurate noise reduction, based on the above-mentioned scheme, the present invention uses a specific DVS as an example, such as Figure 3 As shown in FIG. 1 , according to the detection range of a certain DVS, a storage space W of 128×128×M is set to correspond to each DVS Pixel point one by one. M is the timestamp bit width, which can be adjusted as needed, such as being set to 32.

[0084] However, in other embodiments, the size of the storage space W may be different from the size of the DVS detection plane, and noise reduction may be performed only on a partial area of ​​the DVS detection plane. For example, the DVS detection plane is 256×128, and the size of the storage space W is 128×128×M.

[0085] In a further embodiment, the number of storage levels may be increased to store event information occurring in the storage area W(i, j) before the first event E arrives.

[0086] Step S3: Analyze event attributes and make clustering determination. Specifically, determine the clustering range of the first event in the first storage space according to the coordinate information of the first event; determine whether the first event forms a valid clustering relationship with the events in the clustering range according to the relationship between the time information of the events stored in the clustering range and the time information of the first event, and if so, determine that the first event is a valid event, otherwise, determine that the first event is a noise event.

[0087] In the storage space mapping, all the recent events of DVS at each coordinate are recorded. When the storage space is expanded in chronological order, a series of scattered points can be observed. Here, the clustering relationship between the input event E and the scattered points of each event in the storage space is determined to distinguish between noise and valid events.

[0088] The dynamic events detected by DVS are generally triggered by the edge area of ​​the moving object. The intensity gradient at the edge is large and has a strong correlation in space and time. Therefore, effective events are usually concentrated or continuous points in space. Noise events are often generated randomly. As mentioned above, the noises accompanying DVS imaging include random noise, hot pixel noise, black block noise, black noise, etc.

[0089] For ordinary noise events, such as random noise and / or hot pixel noise, they are usually discretely distributed in space and / or time, appearing as separate and discrete points, with weak correlation with other events, and a small spatial clustering set, which is easy to distinguish from real events. However, for special noise events, such as black block noise and / or black noise, they are not discretely distributed in space and / or time, and have certain temporal and / or spatial correlations, and ordinary noise reduction methods or clustering methods are not sufficient to cope with them. In order to further improve the noise reduction effect, the present invention filters out this type of noise, clusters the area, and distinguishes between valid events and noise events based on the frequency characteristics of events occurring in the area.

[0090] The present invention combines the time and space characteristics of DVS real-time output events, analyzes the first event and the previous DVS detection space events, thereby filtering out DVS noise while ensuring accuracy and retaining valid events, which has important reference and guiding significance for improving DVS imaging performance.

[0091] Furthermore, the present invention distinguishes noise events from valid events by clustering, such as Figure 4 As shown, through clustering analysis, the clustered events are divided into multiple subsets, at least classified into a valid event set and a noise event set, wherein the data in each subset has a high degree of intrinsic similarity. The clustering operation can use K-means clustering or mean shift clustering to achieve effective and real-time classification.

[0092] In addition, during the DVS real-time noise reduction process, for each newly generated event, a clustering operation is performed to determine the clustering relationship between the event and the events within the clustering space in the storage space W, and whether the event is noise or an event is determined based on the clustering relationship. The specific steps are as follows:

[0093] Step S31: Determine the spatial clustering range.

[0094] The spatial clustering range is determined according to the attributes of the input event. For example, the coordinate range associated with the input event in the storage space W can be determined according to the coordinates (i, j) of the input event, such as the range of W(i, j)±d or W(i±d, j±d).

[0095] The clustering range can be set to different values ​​according to the specific conditions such as the DVS sampling speed and the effective event generation speed. For example, if the target moves fast, the effective event generation rate is fast, and the associated coordinate range in the storage space W can be larger. Correspondingly, if the event generation rate is slow, the associated coordinate range in the storage space W can be smaller.

[0096] In addition, the coordinate range associated in the storage space W may also have other representation ranges, such as a coordinate range of a rectangle, a circle or other shapes, which is not limited in the present invention.

[0097] Step S32: Perform cluster analysis to filter out common noise.

[0098] In the storage space W, the events are clustered within the address range associated with the storage unit corresponding to the first event (i.e., the clustering range). Through the clustering operation, it can be determined whether the event forms a valid clustering relationship with the events in the clustering range, and based on the clustering relationship, it can be determined whether the event is a noise event or a valid event.

[0099] The clustering judgment is easy to implement in hardware and can be designed using FPGA, ASIC (such as a dedicated interface circuit) or IC. The calculations and judgments involved are all processed in parallel. Parallel processing is faster and has stronger real-time performance, avoiding the delay and increased power consumption caused by the accumulation operation performed when the software performs noise reduction processing.

[0100] Preferably, a threshold is used to perform clustering judgment to judge whether the first event is a valid event or a noise event.

[0101] like Figure 5 As shown, when the DVS outputs the first event, the computing unit determines the clustering range according to the DVS coordinates and the clustering settings, such as the clustering range is a 17×17 pixel square around the first event. In the storage space, the timestamp of each pixel in the clustering range can be obtained by accessing adjacent addresses.

[0102] According to the above steps, information about the first event and the events within its clustering range can be obtained. This information includes at least time information. Through clustering operations, it is determined whether the event is a noise event or a valid event.

[0103] Here, a clustering operation scheme applied to hardware is briefly introduced. This scheme is implemented in hardware:

[0104] Let the data information of the first event E generated by DVS be: coordinates (i, j), timestamp T(i, j), and the clustering range be [(i - 8, j - 8), (i - 8, j - 8 + 1), …, (i + 8, j + 8)]. Let value(k, p) be the value representing the relationship in terms of time between the first event E(i, j) and an event E(k, p) within the clustering range (referred to as the time relationship value), and the timestamp of event E(k, p) be T(k, p).

[0105] The above settings have achieved spatial screening. Based on the need for simplified operations, only the time difference is considered here to simplify the operations, that is, calculate the time interval (or time difference) dT(k, p) between the first event E(i, j) and each event E(k, p) within the clustering range:

[0106] dT(k, p) = T(k, p) - T(i, j);

[0107] Set a first threshold T1, compare the time interval dT with the first threshold T1 to obtain the time relationship value between the first event E(i, j) and any event E(k, p) within the clustering range. And in some embodiments, the first relationship value is 1 and the second relationship value is 0.

[0108] When dT < T1, value(k, p) = 1 (the first relationship value), representing that the first event E(i, j) is related to an event E(k, p) within the clustering range in terms of time. When dT > T1, value(k, p) = 0 (the second relationship value), representing that they are not related in terms of time.

[0109] The time relationship value in the present invention numerically represents the time correlation based on the time interval between two events. In other embodiments, the relationship value can be other values. For example, the first relationship value can be 2 or any positive number, and the second relationship value can be -1 or any non-positive number. As described above, preferably, when the time interval is greater than a certain range, its relationship value is determined to be 0, otherwise it is a non-zero value.

[0110] Alternatively, any other reasonable pair of values representing proximity (such as 0.01 and 1) can be used as the time relationship value between two events. At the same time, the relationship value can also be set to the opposite values. For example, the relationship value of a valid event is 0, and the relationship value of a noise event is 1 or -1. The present invention does not limit this.

[0111] It is very easy to foresee that the logical judgment and numerical assignment of at least this position in the present invention can be reasonably set / adjusted according to the concept of the present invention to meet the logical needs of solving the noise reduction problem. This deformation of logical judgment and numerical assignment does not deviate from the concept of the present invention, is an equivalent alternative to the present invention, and also belongs to the technical solution claimed to be protected by the present invention.

[0112] The distribution of several events within the cluster that have a temporal relationship with the first event is as follows: Figure 6 As shown in gray, add the time relationship value between the first event and any event within the cluster range to obtain the total cluster time relationship value Value total , then the sum of clustering time relationship values ​​is:

[0113]

[0114] In certain embodiments, the time relationship between the first event and the events within the clustering range may also be processed in a sequential (serial) read / write manner.

[0115] Taking the first relationship value as 1 and the second relationship value as 0 as an example, the total value of the clustering time relationship value is total Set the second threshold: When Value total If the value is greater than the second threshold, it is considered that the first event forms an effective clustering relationship with the events within the clustering range, has a strong correlation, and can be considered to belong to the same event set, that is, dynamic behavior; when Value total If the value is less than the second threshold, it is considered that the first event does not form an effective clustering relationship with the events within the clustering range, has a weak correlation, and can be considered not to belong to the same event set, that is, non-dynamic behavior, and can be considered as noise. Based on this, the real event can be identified and the noise can be filtered out.

[0116] In order to improve the data processing speed, the calculation of the above-mentioned time relationship value or / and the sum of the cluster time relationship values ​​can be configured to be executed in parallel, that is, the time relationship between the first event E(i, j) and any event within the cluster range is calculated in parallel, or / and the calculation and judgment of the sum of the time relationship values ​​are completed.

[0117] Therefore, the calculation and judgment involved in the process of using the first threshold and the second threshold to determine whether the event is a noise event or a valid event can be processed in parallel. Parallel processing is implemented in hardware and occupies less resources, which can effectively reduce the output power consumption of the DVS sensor.

[0118] At the same time, this embodiment uses very simple operations. For example, the time interval between the first event and each historical event within the cluster range is calculated in parallel by addition and / or subtraction operations, and the time interval between the first event and each historical event within the cluster range is compared in parallel with the first threshold. It has the characteristics of real-time and high speed, and avoids serial calculations (such as the accumulation operations required for software noise reduction processing) that increase delays and power consumption.

[0119] Furthermore, the present invention also includes step S33: filtering out special noise.

[0120] For events within the associated coordinate range in the storage space W (events in spatial clustering), whether they are noise is further determined based on the time information of these events. Figure 7 As shown, the timestamp information of events in spatial clustering is concentrated from t1 to t n , calculate the average time interval or maximum time interval of these events. Since the frequency of noise events such as black block noise and black noise is extremely high, often greater than the frequency of events, this type of special noise can be avoided through rejection.

[0121] Preferably, a third threshold is set, and the third threshold is a time threshold or a frequency threshold. When the third threshold is a time threshold, if the time interval (average time interval or / and maximum time interval) of all events in the spatial clustering range is less than the third time threshold, the event is considered to be a noise event, otherwise, the event is considered to be a valid event. When the third threshold is a frequency threshold, if the average frequency or / and the minimum frequency of all events in the clustering range is greater than the third threshold, the event is considered to be a noise event, otherwise it is a valid event.

[0122] The present invention divides noise into common noise events (random noise, hot pixel) and special noise events (black block noise, black noise) according to the temporal and / or spatial correlation of noise events. Common noise events are filtered out according to step S32, and special noise events are filtered out according to step S33. When the two are combined, the influence of noise can be better filtered out. Therefore, joint determination can also be considered, such as Figure 8 The step S33 of reducing the special noise can be placed before step S32 or after step S32, or can be processed in parallel, which is not limited in the present invention.

[0123] Fig. 9It is a simplified diagram of the clustering noise reduction method of the present invention. The real-time noise reduction technology of the present invention mainly adopts the following method: receiving the events generated by the DVS in real time, and continuously storing the first event into the address corresponding to the storage space W, and the storage space W changes dynamically and is updated in real time. The clustering range is determined with the position of the first event in the storage space as the center, and it is determined whether the first event forms an effective clustering relationship with the events within the clustering range. According to the clustering relationship, it can be determined whether the event belongs to noise or event. Further, according to the average time interval or / and the maximum time interval of all events in the clustering space, special noise (black block noise, black noise) is filtered out, wherein this step can be placed before filtering out the ordinary noise event set, or it can be placed after filtering out the ordinary noise event set, or it can be processed in parallel.

[0124] In addition, through cluster analysis, clustering judgment is performed on events stored within a certain range (clustering range) in the storage space, and the events within the clustering range are classified into at least valid events and noise events, thereby obtaining a valid event set and a noise event set. The valid event set is retained, and the noise event set can be retained for the judgment of the subsequent first event, or the noise event set can be discarded or filtered out. However, for the method of discarding the noise event set, this method has the defect of misjudgment, that is, some valid events are misjudged as noise events, resulting in the loss of some valid events. For example, when the target moves, the first few valid events collected by the DVS are likely to be misjudged as noise events.

[0125] The clustering real-time noise reduction technology of the present invention is adopted, and the sum of the clustering time relationship values ​​is 6 as the judgment basis, that is, the second threshold is set to 6, and the filtering effect is as follows: Fig.10a and Fig.10b , Fig.11a and Fig.11b , Fig.12a and Figure 12b shown.

[0126] Fig.10a and Fig.10b This is a comparison chart of the noise reduction effect of black noise in a certain DVS, where: Fig.10a This is the black noise spectrum of DVS before noise reduction. Fig.10b It represents the black noise spectrum after adopting the noise reduction technology of the present invention. Fig.11a and Fig.11b This is a comparison chart of the noise reduction effect of a certain DVS on the black block noise caused by the black block object when there is a black block object in the field of view. Fig.11a This is the black block noise map caused by the black block object before noise reduction. That is, when the DVS is in a static state, the black block object in the field of view generates a lot of noise. Fig.11b It represents the black block noise spectrum after adopting the noise reduction technology of the present invention.

[0127] It can be seen that by using the clustering real-time noise reduction technology of the present invention, the black noise and black block noise existing in the DVS imaging are basically eliminated after noise reduction, and the unexpected environmental noise is effectively filtered out, that is, when there is no moving object in the DVS field of view, the first event is not generated, which effectively reduces the system power consumption.

[0128] Fig.12a and Figure 12b This is a comparison diagram of the imaging effect of a certain DVS before and after the noise reduction technology of the present invention is used when there is a moving object in the field of view of a certain DVS (activated state). Fig.12a It shows the event spectrum (including noise events and valid events) when DVS is activated before filtering. Figure 12b The event map after the noise reduction technology of the present invention is used when the DVS is activated can be shown. It can be seen that the clustering real-time noise reduction technology of the present invention can effectively filter out noise events, and the interference of noise events is almost non-existent. The motion trajectory is lossless and clearly visible, and the thickness of the entire motion trajectory is basically consistent, which intuitively shows that the noise reduction process has little impact on effective events and a good noise reduction effect. Therefore, the present invention can filter out multiple noises while ensuring accuracy, can maximize the retention of effective events, and has little damage to effective events, which has important reference and guiding significance for improving the performance of event imaging devices.

[0129] The other undescribed contents of this type of embodiment are the same as those of the previous embodiment. To avoid the text being too lengthy, all the contents of the previous embodiment are cited here by reference.

[0130] The pulse neural network processor or event-driven processor is very compatible with the event stream output by the event imaging device. They are all based on asynchronous processing of pulse events, so that only a small amount of information is transmitted in the system, achieving low power consumption and real-time performance. Fig.13 As shown, the configuration parameters such as the trained weight values ​​are deployed in the brain-like chip, and the pulse neural network in the chip performs reasoning based on the denoised event stream, thereby improving the accuracy of the reasoning or classification results of the brain-like chip.

[0131] In addition, the clustering noise reduction technology of the present invention can be integrated into the event imaging device, arranged in the processor or / and configured in the interface between the event imaging device and the processor, wherein the interface circuit can exist independently or be integrated with the pulse neural network processor or the event-driven processor on a single chip. The integration can be coupled to the single chip through an adapter board or manufactured in the same bare die.

[0132] The present invention also provides a chip, which includes the clustering noise reduction device as described in any one of the above, or is configured to perform the above steps S1 to S3 and all other steps. The processor / chip includes but is not limited to a pulse neural network processor, an event-driven processor, a brain-like chip, a brain-like processor, etc.

[0133] The present invention also provides an event imaging device, which includes the clustering noise reduction device as described in any one of the above items, or the event imaging device applies the clustering noise reduction method as described in any one of the above items. That is, the noise reduction scheme of the present invention can be directly applied to the event imaging device, and the two are integrated into the same product.

[0134] The present invention also includes an electronic device, which includes an event imaging device and a clustering noise reduction device as described in any of the above items, wherein the clustering noise reduction device is configured to filter noise on an event stream output by the event imaging device.

[0135] In order to achieve better technical effects or meet the needs of certain applications, those skilled in the art may make further improvements to the technical solution based on the present invention. However, even if such partial improvement / design is creative and / or progressive, as long as the technical features covered by the claims of the present invention are utilized, the technical solution should also fall within the protection scope of the present invention according to the "comprehensive coverage principle".

[0136] Some of the technical features mentioned in the attached claims may have alternative technical features, or the order of certain technical processes and the order of material organization may be reorganized. After knowing the present invention, ordinary technicians in this field can easily think of these replacement means, or change the order of technical processes and the order of material organization, and then adopt basically the same means to solve basically the same technical problems and achieve basically the same technical effects. Therefore, even if the above-mentioned means and / or order are clearly defined in the claims, these modifications, changes, and substitutions should fall within the scope of protection of the claims based on the "principle of equivalents".

[0137] For the numerical values ​​clearly defined in the claims, those skilled in the art will generally understand that other reasonable values ​​near the numerical values ​​can also be applied to a specific implementation. These design solutions that circumvent the concept of the present invention by means of details also fall within the scope of protection of the claims.

[0138] In conjunction with the various method steps and units described in the embodiments disclosed herein, they can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the steps and components of each embodiment have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of protection claimed by the present invention.

Claims

1. A clustering noise reduction device, the clustering noise reduction device receiving a plurality of events output by an event imaging device, wherein the data information of the plurality of events at least includes coordinate information and time information, and the plurality of events includes a first event, characterized in that: The clustering noise reduction device comprises: The first storage space is configured to: at least store time information of an event; A mapping module is configured to: store at least time information of the first event in a first storage unit located in a first storage space and corresponding to the coordinate information of the first event according to the coordinate information of the first event; The clustering operation module is configured to: determine the clustering range of the first event in the first storage space according to the coordinate information of the first event; determine whether the first event forms a valid clustering relationship with the events in the clustering range according to the relationship between the time information of the events stored in the clustering range and the time information of the first event, and if so, determine that the first event is a valid event, otherwise determine that the first event is a noise event.

2. The clustering noise reduction device according to claim 1, characterized in that: After at least time information of some events generated later than the first event among the plurality of events is stored in the first storage space, determining whether the first event forms a valid clustering relationship with the events within the clustering range according to the relationship between the time information of the events stored within the clustering range and the time information of the first event; or, Before or after at least the time information of the first event is stored in the first storage unit located in the first storage space and corresponding to the coordinate information of the first event, it is immediately determined whether the first event forms a valid clustering relationship with the events within the clustering range based on the relationship between the time information of the events stored within the clustering range and the time information of the first event.

3. The clustering noise reduction device according to claim 1, characterized in that: The clustering range of the first event is: in the event imaging device, an address range in the first storage space in which event imaging units that are physically adjacent to the event imaging unit that generates the first event are mapped.

4. The clustering noise reduction device according to any one of claims 1 to 3, characterized in that: The clustering operation module includes a first threshold judgment module and a second threshold judgment module; The first threshold determination module is configured to: calculate the time interval between the first event and each event within the cluster range based on the time information of the event, compare the calculated time interval with the first threshold, and obtain the time relationship value between the first event and each event within the cluster range; Add the time relationship value between the first event and each event within the cluster range to obtain the sum of the cluster time relationship values; The second threshold judgment module is configured to compare the sum of the clustering time relationship values ​​with a second threshold, and determine whether the event is a noise event or a valid event according to the comparison result.

5. The clustering noise reduction device according to any one of claims 1 to 3, characterized in that: The clustering operation module includes a third threshold judgment module; The third threshold judgment module is provided with a third threshold, wherein the third threshold is a time threshold or a frequency threshold; If the third threshold is a time threshold, if the average time interval and / or the maximum time interval of all events within the cluster range is less than the third threshold, the event is determined to be a noise event; If the third threshold is a frequency threshold, if the average frequency and / or the minimum frequency of all events within the clustering range is greater than the third threshold, the event is determined to be a noise event.

6. The clustering noise reduction device according to claim 4, characterized in that: The first threshold determination module: calculates the time relationship value between the first event and each event within the cluster range in parallel; or / and, adds the time relationship value between the first event and each event within the cluster range in parallel.

7. The clustering noise reduction device according to claim 4, characterized in that: If the time interval is greater than the first threshold, the time relationship value is 0; if the time interval is less than the first threshold, the time relationship value is 1; if the sum of the cluster time relationship values ​​is greater than the second threshold, it is determined to be a valid event, otherwise it is a noise event.

8. A clustering denoising method, characterized in that: The clustering denoising method comprises the following steps: Receiving a plurality of events output by an event imaging device, wherein data information of the plurality of events at least includes coordinate information and time information, and the plurality of events includes a first event; According to the coordinate information of the first event, storing at least the time information of the first event in a first storage unit located in the first storage space and corresponding to the coordinate information of the first event; Determining a clustering range of the first event in a first storage space according to the coordinate information of the first event; According to the relationship between the time information of the events stored within the clustering range and the time information of the first event, it is determined whether the first event forms a valid clustering relationship with the events within the clustering range. If so, the first event is determined to be a valid event; otherwise, the first event is determined to be a noise event.

9. The clustering denoising method according to claim 8, characterized in that: After at least time information of some events generated later than the first event among the plurality of events is stored in the first storage space, determining whether the first event forms a valid clustering relationship with the events within the clustering range according to the relationship between the time information of the events stored within the clustering range and the time information of the first event; or, Before or after at least the time information of the first event is stored in the first storage unit located in the first storage space and corresponding to the coordinate information of the first event, it is immediately determined whether the first event forms a valid clustering relationship with the events within the clustering range based on the relationship between the time information of the events stored within the clustering range and the time information of the first event.

10. The clustering denoising method according to claim 8, characterized in that: The clustering range of the first event is: in the event imaging device, an address range in the first storage space in which event imaging units that are physically adjacent to the event imaging unit that generates the first event are mapped.

11. The clustering denoising method according to any one of claims 8 to 10, characterized in that: Based on the time information of the event, respectively calculate the time interval between the first event and each event within the cluster range, compare the calculated time interval with the first threshold, and obtain the time relationship value between the first event and each event within the cluster range; Add the time relationship value between the first event and each event within the cluster range to obtain the sum of the cluster time relationship values; The sum of the clustering time relationship values ​​is compared with a second threshold, and the event is determined to be a noise event or a valid event according to the comparison result.

12. The clustering denoising method according to any one of claims 8 to 10, characterized in that: Setting a third threshold, wherein the third threshold is a time threshold or a frequency threshold; If the third threshold is a time threshold, if the average time interval and / or the maximum time interval of all events within the cluster range is less than the third threshold, the event is determined to be a noise event; If the third threshold is a frequency threshold, if the average frequency and / or the minimum frequency of all events within the clustering range is greater than the third threshold, the event is determined to be a noise event.

13. The clustering denoising method according to claim 11, characterized in that: Calculate the time relationship value between the first event and each event within the cluster in parallel; or / and, add the time relationship value between the first event and each event within the cluster in parallel.

14. The clustering denoising method according to claim 11, characterized in that: If the time interval is greater than the first threshold, the time relationship value is 0; if the time interval is less than the first threshold, the time relationship value is 1; if the sum of the cluster time relationship values ​​is greater than the second threshold, it is determined to be a valid event, otherwise it is a noise event.

15. A clustering noise reduction method, the clustering noise reduction method receiving a first event output by an event imaging device, characterized in that: Acquire data information of the first event, the data information at least including coordinate information; wherein the coordinate information is the position of the first event when the detection plane of the event imaging device is generated; According to the coordinate information, mapping the first event to a corresponding first storage unit in the first storage space; Determining a clustering range of a first event in a first storage space; Through clustering operation, the events within the clustering range are classified to obtain at least a valid event set and a noise event set.

16. The clustering denoising method according to claim 15, characterized in that: The clustering operation adopts K-means clustering and / or mean shift clustering.

17. A chip, characterized in that: The chip includes the clustering noise reduction device as described in any one of claims 1 to 7, or the chip applies the clustering noise reduction method as described in any one of claims 8 to 16.

18. The chip according to claim 17, characterized in that: The clustering noise reduction device is implemented in a dedicated interface circuit; or implemented as an FPGA circuit; or implemented by combining software and hardware; or the clustering noise reduction method is implemented in a host computer through software.

19. An event imaging device, characterized in that: The event imaging device includes the clustering denoising device as described in any one of claims 1-7, or the event imaging device applies the clustering denoising method as described in any one of claims 8-16.

20. An electronic device, characterized in that: The electronic device comprises an event imaging device and a clustering noise reduction device as described in any one of claims 1 to 7, wherein the clustering noise reduction device is configured to perform noise filtering on a plurality of events output by the event imaging device.

Images