Biomimetic moving target sensing, storage and calculation integrated photoelectric detection system based on continuous photoconduction

Abstract

The invention discloses a bionic moving target sensing, storage and calculation integrated photoelectric detection system based on continuous photoconduction, and relates to the field of bionic intelligent vision and the field of brain-like calculation, and the system comprises a synaptic bionic photoelectric detector array and a brain-like memristor array which are interconnected. When the synaptic-like bionic photoelectric detector array detects visual information of a moving target, a fixed bias voltage is applied to a device in each pixel, space illumination information corresponding to each pixel is converted into a continuously attenuated photocurrent signal, illumination information of a multi-frame dynamic target is continuously received in the continuous attenuation process of the photocurrent, and the visual information of the moving target is detected by the synaptic-like bionic photoelectric detector array. After the last frame of illumination information is input, a light current signal of each pixel is read, and a feature map memorizing all historical moment information is determined; the brain-like memristor array is based on the Ohm's law and the Kirchhoff's law, performs deep neural network operation on the feature map, outputs a current corresponding to a task result label, identifies a dynamic target, and can meet the requirements of a deep neural network and a complex dynamic visual task.

Description

Technical Field

[0001] The present application relates to the fields of bionic intelligent vision and brain-like computing, and in particular to a bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity. Background Art

[0002] In recent years, with the rapid development of artificial intelligence applications, many new application scenarios have emerged, which has put forward higher requirements for detection systems. Traditional detection systems usually adopt discrete architectures, and imaging calculations face high latency, which makes it difficult to meet the needs of fast detection, recognition, and judgment in fields such as intelligent driving and robot vision. Therefore, it is urgent to develop lightweight and intelligent infrared photoelectric perception systems with edge computing capabilities to achieve fast detection and recognition of targets.

[0003] In dynamic motion recognition tasks, since objects are constantly moving, the detector needs to perform frequent frame-by-frame analysis to ensure that the motion trajectory and dynamic changes can be accurately captured and identified. This frame-by-frame analysis paradigm has led to a sharp increase in data transmission volume, and has placed extremely high demands on data transmission efficiency.

[0004] At present, the discrete architecture of detection and processing modules in traditional visual perception systems results in the relatively simple functions of bionic sensing, storage and computing integrated photoelectric detection systems. They can only build simple single-layer perception models and implement basic image preprocessing tasks, making it difficult to meet the needs of deep neural networks and complex dynamic visual tasks. Summary of the invention

[0005] The purpose of this application is to provide a bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity to solve the problem that existing bionic sensing, storage and computing integrated photoelectric detection systems are difficult to cope with the requirements of deep neural networks and complex dynamic visual tasks.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity, comprising: an interconnected synaptic-like bionic photoelectric detector array and a brain-like memristor array.

[0008] When the synaptic-like bionic photodetector array is used to detect visual information of a moving target, a fixed bias voltage is applied to the device in each pixel in the synaptic-like bionic photodetector array to convert the spatial illumination information corresponding to each pixel into a continuously decaying photocurrent signal. During the continuous decay of the photocurrent, multiple frames of illumination information of dynamic targets are continuously received, and after the last frame of illumination information is input, the photocurrent signal of each pixel is read to determine a feature map that memorizes information of all historical moments.

[0009] The brain-like memristor array is used to perform deep neural network operations on the feature map based on Ohm's law and Kirchhoff's law, output a current corresponding to the task result label, and identify dynamic targets based on the current; the current is used to characterize the probability of the task result label.

[0010] According to the specific embodiments provided in this application, this application discloses the following technical effects.

[0011] When the present application uses a synaptic-like bionic photodetector array to detect visual information of moving targets, a fixed bias is applied to the device in each pixel to convert the spatial illumination information corresponding to each pixel into a continuously decaying photocurrent, and in the process of continuous decay of the photocurrent, illumination information of multiple frames of dynamic targets is continuously received without reading out the electrical signal frame by frame. The magnitude of the photocurrent decays or increases continuously over time, and the photocurrent state of each pixel of the array is read out after the last frame of light information is input to obtain a feature map that memorizes all historical moment information, thereby realizing multi-frame fusion detection calculation in the intra-sensory time domain; the photocurrents extracted after all pixel detections are combined into a feature map as the output of the first-layer network, which is input into the brain-like memristor array, and a deep neural network operation is performed. Finally, a current corresponding to the task result label is output to represent its probability, thereby completing the recognition and classification tasks of dynamic targets; a storage-computation integrated deep neural network operation is performed in the memristor array, the conductance value of each memristor in the memristor array is read by applying a voltage sequence, the output current of each memristor is calculated, and the output current is used as the result of the recognition task. Among them, the synaptic bionic photoelectric detector array utilizes the continuous photoconductivity effect of the detector when performing multi-frame fusion detection and reading in the time domain. The photocurrent decays nonlinearly with time, and the linear and nonlinear calculations of the single-layer neural network are completed at the same time; the brain-like memristor array realizes fully connected neural network calculations through Ohm's law and Kirchhoff's law, outputs current results, and completes the dynamic target classification task.

[0012] Compared with the traditional visual perception system's discrete architecture of detection and processing modules and the frame-by-frame transmission computing paradigm, this application can meet the needs of deep neural networks and complex dynamic visual tasks, greatly improve the network's recognition speed, reduce network power consumption, and achieve rapid recognition of moving targets at the full hardware level. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0014] Figure 1 This is a schematic diagram of the structure of the retina-like photodetector unit of the present application.

[0015] Figure 2 This is a schematic diagram of the structure of the positive electrode and the negative electrode of the retinal photodetector unit of the present application.

[0016] Figure 3 This is a schematic diagram of a 1T1R memristor array used in this application.

[0017] Figure 4 This is the IV characteristic curve of the retina-like photodetector unit of the present application.

[0018] Figure 5 This is a continuous photoconductivity characteristic curve of the retinal photodetector unit of the present application.

[0019] Figure 6 This is a schematic diagram of the structure of a 4×4 pixel synaptic-type bionic photodetector array of the present application.

[0020] Figure 7 This is a schematic diagram of the bionic target sensing, storage and computing integrated photoelectric detection system of this application.

[0021] Figure 8 A heat map of the conductance values ​​of the memristor crossbar array of the present application.

[0022] Fig. 9 This is a curve chart of recognition accuracy for the dynamic target recognition task scenario of this application. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0024] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0025] In order to realize large-scale integrated sensing, storage and computing devices and support the processing of deep neural networks and complex image recognition tasks, it is urgent to break through network integration technology and promote the joint design of synaptic-like bionic photodetectors and brain-like neural network hardware. Memristors, with their brain-like synaptic functions and excellent integration capabilities, provide new opportunities for integrated sensing, storage and computing technology. However, there is currently no research on bionic intelligent photodetector arrays based on continuous photoconductivity multi-frame detection and their integrated integration with brain-like memristor arrays.

[0026] In order to meet the high requirements for data transmission efficiency, improving the processing power inside the detector has become the key to solving the energy consumption and delay problems in machine vision systems. Based on the research on some specific materials and device structures, the applicant found that detector materials and structures with continuous photoconductivity effect can provide a good idea for solving this bottleneck. In this type of device, whenever the illumination ends, the photocurrent decays over time instead of disappearing instantly. By utilizing the continuous photoconductivity effect of the detector, the present application proposes a new paradigm for moving target recognition based on multi-frame unified imaging calculation. In the process of continuous attenuation of the photocurrent, multiple frames of light signals are continuously detected, and the continuous multi-frame image information is stored in a single frame for reading. Without affecting the accuracy, the number of data transmission and processing times is reduced exponentially, which is very beneficial to reducing power consumption and delay.

[0027] Based on this, an embodiment of the present application provides a bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity, including: an interconnected synaptic-type bionic photoelectric detector array and a brain-like memristor array.

[0028] When the synaptic-like bionic photodetector array is used to detect visual information of a moving target, a fixed bias voltage is applied to the device in each pixel in the synaptic-like bionic photodetector array to convert the spatial illumination information corresponding to each pixel into a continuously decaying photocurrent signal. During the continuous decay of the photocurrent, multiple frames of illumination information of dynamic targets are continuously received, and after the last frame of illumination information is input, the photocurrent signal of each pixel is read to determine a feature map that memorizes information of all historical moments.

[0029] The brain-like memristor array is used to perform deep neural network operations on the feature map based on Ohm's law and Kirchhoff's law, output a current corresponding to the task result label, and identify dynamic targets based on the current; the current is used to characterize the probability of the task result label.

[0030] In an exemplary embodiment, the synapse-like bionic photodetector array specifically includes: a plurality of pixels distributed in an array, each pixel including a retina-like photodetector unit.

[0031] In practical applications, the synaptic-like bionic photodetector array includes 4×4 pixels, each pixel contains one retinal-like photodetector unit, and a total of 16 retinal-like photodetector units.

[0032] In an exemplary embodiment, each of the retinal-like photodetector units includes a substrate, a light absorption layer and two crossed interdigitated electrodes stacked in sequence from bottom to top; wherein the light absorption layer region corresponding to the channel formed between the two crossed interdigitated electrodes is a light response region; the light response region is used to provide linear light response and nonlinear attenuation of photocurrent, and to regulate the voltage of the electrodes at both ends of each of the retinal-like photodetector units to encode the light response of the retinal-like photodetector unit.

[0033] In an exemplary embodiment, the substrate is an insulating Al2O3 substrate; the light absorption layer is a MoS2 two-dimensional material; an image design is performed on the MoS2 two-dimensional material by a photolithography process, and a metal electrode is deposited on the surface of the MoS2 two-dimensional material by electron beam evaporation; the metal electrode is the interdigitated electrode.

[0034] In an exemplary embodiment, the interdigitated electrodes are rectangular cross-finger structures, forming a plurality of light response units; the two interdigitated electrodes include a positive electrode and a negative electrode; the positive electrode and the negative electrode are arranged in an alternating manner.

[0035] In practical applications, the width of the positive electrode and the negative electrode is 15 μm, and the spacing between adjacent electrodes is 5 μm.

[0036] In an exemplary embodiment, the photoresponsivity of the retina-like photodetector unit is a random fixed value; The photocurrent signal is the product of the photoresponse rate and the light intensity.

[0037] In an exemplary embodiment, the photocurrent signal exhibits a nonlinear continuous decay between frames.

[0038] In practical applications, the synaptic bionic photodetector array realizes the calculation of a single-layer nonlinear recurrent neural network. The light response rate R of the synaptic bionic photodetector array represents the network weight value. The light intensity P is used as the input signal of the network. The generated photocurrent signal is .

[0039] The photocurrent continues to decay after it is generated, and during the process of continuous decay of the photocurrent, multiple frames of dynamic target illumination are continuously received. The photocurrent signal at the next moment is the remaining photocurrent signal after the last moment And the light intensity generated by the light at the next moment Joint decision, i.e. .

[0040] In an exemplary embodiment, the brain-like memristor array utilizes word line circuits, source line circuits, and bit line circuits interwoven into a mesh layout; a memristor unit is arranged at each intersection in the mesh layout; and these three circuits are used to input or read signals in the memristor unit.

[0041] The memristor unit includes a transistor and a memristor.

[0042] The word line circuit is connected to the gate of the transistor, the source line circuit is connected to the top electrode of the memristor, and the bit line circuit is connected to the bit of the memristor.

[0043] The bottom electrode of the memristor is also directly connected to the drain of the transistor; the conductive state of the brain-like memristor array is effectively controlled by adjusting the gate voltage of the transistor; the conductive state includes an on state and an off state. This design can effectively solve the signal crosstalk problem and improve the performance and stability of the brain-like memristor array.

[0044] In practical applications, each memristor in the brain-like memristor array can adjust its resistance value according to the applied voltage, and keep the conductance value unchanged after the external voltage is removed, and can stably maintain the adjusted resistance state, so as to deploy the weights of the hardware deep neural network, which is the brain-like memristor array.

[0045] As the hardware foundation for neural network implementation, these memristors can accurately map and write the hardware deep neural network weights obtained through training into their respective resistance values ​​by applying the corresponding voltage.

[0046] In an exemplary embodiment, the top layer of the memristor array is a top electrode layer, the middle layer is a resistive switching layer, and the bottom layer is a bottom electrode layer.

[0047] The top electrode layer is input with a voltage through row selection, and the bottom electrode layer is output with a current through column selection.

[0048] The memristor array follows Ohm's law and Kirchhoff's law to complete multiplication and accumulation operations, reads current from each column, completes vector-matrix multiplication operations in the deep neural network, and realizes a hardware deep neural network with integrated storage and computing.

[0049] In practical applications, the brain-like memristor array performs storage-computing integrated deep neural network operations, and the extracted feature maps are input into the hardware deep neural network for recognition. is the voltage value expanded by each pixel of the image information, obtained by multiplying the photocurrent obtained by the synaptic bionic photodetector array by a fixed coefficient. is the conductance of each memristor, according to Ohm's law and Kirchhoff's law , performs storage-computation integrated deep neural network operations, and finally outputs current Represents the task identification result.

[0050] In an exemplary embodiment, it also includes: a driving circuit; the driving circuit is directly connected to each electrode of the synaptic-type bionic photodetector array to read the photocurrent signal.

[0051] The driving circuit is directly connected to the top electrode of the brain-like memristor array, and is used to convert the photocurrent signal into a voltage signal, and input the voltage signal into the brain-like memristor array.

[0052] The driving circuit includes a power supply module, a bias control module, a signal reading module and a host computer, and is used to provide necessary auxiliary functions of power supply, reading, and current-to-voltage conversion.

[0053] This application realizes the intrasensory computing and readout of a synaptic-like bionic photodetector array by constructing a collaborative design of sensitivity control and signal reading.

[0054] This application uses a new computing paradigm of multi-frame integration of sensing, storage and computing to perceive and identify moving targets, simulates the biological functions of human retinal receptors through a synaptic-like bionic photodetector array, and uses the calculation formula of photocurrent signals and the recurrent neural network computational paradigm , nonlinear iterative calculations are performed at the moving target detection end, all characteristic information of the entire motion history process is retained in a characteristic map, and then the current is converted into voltage and directly input into the brain-like memristor storage and computing neural network array to complete the recognition task. The memristor-based storage and computing neural network, that is, the brain-like memristor array, can realize the direct calculation and storage of neural network weight information, thereby eliminating the energy consumption required for data exchange between computing and storage units. Through Ohm's law and Kirchhoff's law , performing storage-computation fusion deep neural network operations on the brain-like memristor array, efficiently completing the infrared image recognition task.

[0055] In another exemplary embodiment, inspired by the biological visual system, the present application is based on a neuromorphic photoelectric detector, which fuses multiple frames of image information of a moving target into one frame and reads it out as a feature map, which is input into a hardware deep neural network with brain-like functions for recognition. In addition, each pixel integrates dual detector units to achieve multi-feature extraction and wide dynamic range detection, and ultimately develops a bionic target sensing, storage and computing integrated photoelectric detection system with fast edge computing capabilities.

[0056] The retina-like photodetector unit in the present application, hereinafter referred to as the device, adopts the MSM structure to achieve linear light response.

[0057] like Figure 1-Figure 2 As shown, the device and array are based on a molybdenum disulfide (MoS2) light absorption layer 2 on a sapphire substrate 1. A positive electrode 3 is led out at one end of each device, and a negative electrode 4 is led out at the other end. During detection, a fixed bias voltage is applied between the positive electrode 3 and the negative electrode 4, and the current size of the negative electrode 4 is recorded.

[0058] The implementation method is as follows: a MoS2 light absorption layer 2 of about 3-5 atomic layers is grown by chemical vapor deposition on an insulating sapphire (aluminum oxide Al2O3) substrate 1, a graphic design is performed on the MoS2 light absorption layer 2 by photolithography and etching technology, and then a metal electrode is deposited on the surface by electron beam evaporation; finally, the mask is washed, and the photogenerated carriers generated by the device under light are conducted out through the positive electrode 3 and the negative electrode 4.

[0059] In the present application, the MSM structure is a metal-semiconductor-metal structure.

[0060] The synapse-like bionic photodetector array in the present application is composed of 4×4 16 retina-like photodetector units, and the photocurrent signal of the detector is read through a designed signal reading module.

[0061] In this application, a synaptic-like bionic photodetector array is provided with a driving circuit system, which includes a power module, a bias control module, a signal reading module and a host computer. The driving module mainly provides the necessary auxiliary functions of power supply, readout, and current-to-voltage conversion, and is directly connected to each electrode of the detector array through a circuit. By constructing a collaborative design of sensitivity control and signal reading, the in-sensitivity calculation and readout of the detector array are realized.

[0062] In this application, a brain-like memristor array is used, and a storage and computing integrated hardware deep neural network is built through the cross array. Each memristor can adjust the resistance value through an external voltage, and continue to maintain the current resistance value after the external voltage is removed. Therefore, it is used as the hardware carrier of the neural network, and the trained neural network weights are mapped to resistance values ​​and written into the memristor through an external voltage.

[0063] Each memristor unit adopts a 1T1R structure, which refers to a structure combining a transistor and a memristor, where "1T" represents a transistor, which is a field effect transistor, and "1R" represents a memristor. In the 1T1R structure, when a certain unit is operated, the corresponding transistor is turned on, while the corresponding transistors of other units are turned off, thus avoiding the problem of misoperation of surrounding units and the generation of read crosstalk.

[0064] like Figure 3 As shown, the line connecting the gates of a row of transistors is called a word line (WL), the line connecting the top electrodes of a column of brain-like memristors is called a bit line (BL), and the line connecting the sources of a column of transistors is called a source line (SL). When in use, the WL is controlled by an analog switch to turn on the transistors of the specified row, the input signal is added to the BL terminal, and the calculation results of the brain-like memristor array are read from the SL terminal. WL refers to the gates connected to the transistors of each row, which is used to control the switching state of the transistors. SL refers to the source connected to each row of transistors, which is used to provide a path for current. BL refers to the upper electrodes connected to each column of memristors, which is used to read or write data. MUX is a multiplexer circuit used to select a specific row or column, thereby achieving access to a specific memristor unit (or a group of units) in the memristor array.

[0065] In an exemplary embodiment, the present application uses chemical vapor deposition (CVD) to grow MoS2 light absorption layer material, grows a 3nm thick MoS2 light absorption layer on an 800μm thick Al2O3 substrate 1, evenly applies AZ5214 photoresist on the material surface, uses RZX3080 developer for development after exposure, and then sequentially electron beam evaporates 51nmCr and 35nmAu as metal electrodes to form device and array patterns. The graphic structure is drawn using L-Edit software, such as Figure 1-Figure 2 As shown, the device is a rectangular cross-finger structure, the electrode width 5 and the electrode spacing 6 are 15um and 5um respectively, and the MoS2 light absorption layer area between the interdigitated electrodes is the light response area.

[0066] Use acetone stripping, ethanol cleaning, and nitrogen drying; secondly, evenly apply AZ5214 glue on the surface of the material for photolithography again, use RZX3080 developer after exposure to develop so that the photoresist only covers one device in each pixel, and then use 50W oxygen plasma to treat the devices not covered by the photoresist for 3 seconds; finally, use acetone stripping, ethanol cleaning, and nitrogen drying to complete the preparation of the synaptic-type bionic photodetector array.

[0067] Figure 4 The IV characteristic curve diagram of the retina-like photodetector unit of the present application is respectively the IV curve in the dark state and under irradiation of 520nm laser with different powers, which reflects the good light response characteristics of the device. Figure 5This is the continuous photoconductivity characteristic curve of the retinal photodetector unit of the present application. A fixed bias voltage of 0.2 V is applied to the positive and negative electrodes of the device, and three frames of light pulse signals are continuously irradiated with a period of 2s and a pulse width of 100ms. It can be seen that after three consecutive cycles, the photocurrent decays to a final value. This photocurrent value memorizes the historical information of all past moments, proving the good continuous photoconductivity characteristics.

[0068] The devices prepared above were spot welded to a 4×4 array board, as shown in Figure 6 As shown, they are then connected to the signal reading module through DuPont connecting lines.

[0069] The present application adopts lithography, etching, deposition and sputtering techniques to prepare a brain-like memristor array device with a sandwich structure on a silicon substrate.

[0070] First, a top-gate field-effect transistor is prepared on a silicon substrate as a gate device. Then, the bottom electrode region of the memristor is etched out in the drain region of the field-effect transistor through photolithography and etching processes. The metal bottom electrode of the memristor is evaporated in this region through thermal evaporation, and then peeled off under heating conditions to obtain the metal bottom electrode. Then, a metal oxide film is prepared on the bottom electrode through sputtering and deposition processes as the resistive switching layer of the memristor. The oxide materials used are hafnium dioxide HfO2 and tantalum oxide TaO2, which are compatible with the (Complementary Metal-Oxide-Semiconductor, Complementary Metal-Oxide-Semiconductor, CMOS) process. X and other materials; using photolithography and etching processes, the memristor top electrode region is etched on the resistive layer film; then, a metal top electrode is evaporated in this region through a thermal evaporation process, and then peeled off under heating conditions to obtain a metal top electrode. Finally, a memristor cross array with field effect transistors as gates and a 1T1R unit structure is completed, namely a brain-like memristor array, such as Figure 3 shown.

[0071] This application also proposes a precise modulation method based on dynamic pulses for weight adjustment of brain-like memristor arrays to achieve high-precision memristor read and write operations. Each memristor can be precisely controlled by selecting rows and columns in the brain-like memristor array through external control signals. At the same time, combined with time-division multiplexing technology, the demand for digital-to-analog conversion is further reduced, the circuit area is effectively reduced, and the overall power consumption of the circuit is significantly reduced. In the read operation of the brain-like memristor array, the analog switch is first controlled by the WL terminal signal to select the target memristor, and the analog signal is output at the BL terminal through digital-to-analog conversion. At this time, a pulse lower than the threshold voltage (0.2V) is applied to the selected memristor. After the voltage pulse is applied to the memristor, based on Ohm's law, the calculation result is directly obtained, and the output current is read through the SL terminal, and then the resistance value of the current memristor is calculated to realize the reading of the resistance value.

[0072] During the write operation, it is divided into two processes: setting and resetting. The setting process reduces the resistance of the memristor through a positive voltage, while the reset process increases its resistance through a negative voltage. In the setting operation, the analog switch is first controlled by the WL terminal to select the specified memristor, and a pulse higher than the memristor write voltage threshold (1V) is output at the BL terminal through digital-to-analog conversion. After each write pulse, a read pulse is immediately performed to monitor in real time whether the resistance of the memristor reaches the set value. If the preset value is not reached, the voltage of the write pulse is increased by 0.2V until the target resistance is reached; if the resistance of the memristor is lower than the set value after applying the write pulse, a negative voltage pulse is applied until the target value is reached. If the resistance of the memristor is already lower than the target value, a positive voltage pulse is applied until the resistance returns to the set value. In the above manner, the present application realizes the precise adjustment of the memristor weights, and provides an effective technical means for constructing a hardware deep neural network with integrated storage and computing.

[0073] The hardware visual neural network is deployed using the prepared 4×4 pixel synaptic bionic photodetector array and 1T1R brain-like memristor array to achieve multi-frame rapid detection and recognition of dynamic targets.

[0074] The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity is as follows: Figure 7 As shown, the moving target light source continuously illuminates the synaptic bionic photodetector array. Compared with the prior art, the multi-frame integrated sensing, storage and computing new computing paradigm of this application does not read out the photocurrent in each frame. By simulating the biological function of the retinal receptors of the human eye, the photocurrent calculation formula is adopted. and the recurrent neural network computational paradigm , the photocurrent is exposed to multiple illuminations while continuously decaying, and nonlinear iterative operations are performed at the moving target detection end, retaining all characteristic information of the entire motion history process in a characteristic map, and then the current is converted into voltage and directly input into the brain-like memristor storage and computing neural network array to complete the recognition task. The memristor-based storage and computing neural network can realize the direct calculation and storage of neural network weight information, and the trainable neural network weight is mapped to the conductance value of the memristor, such as Figure 8 As shown, Figure 8 This is a heat map of the conductivity value of the memristor crossbar array of the present application, thereby eliminating the energy consumption required for data exchange between computing and storage units. , executing deep neural network operations with storage and computing fusion on the brain-like memristor hardware platform, efficiently completing the infrared image recognition task.

[0075] Fig. 9 This is a recognition accuracy curve diagram of the dynamic target recognition task scenario of the present application. Compared with the traditional network, the present application can complete tasks that the traditional single-layer fully connected neural network technology cannot complete based on single-frame image information.

[0076] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity, characterized in that: The photoelectric detection system comprises: an interconnected synaptic-like bionic photoelectric detector array and a brain-like memristor array; When the synaptic-like bionic photodetector array is used to detect visual information of a moving target, a fixed bias voltage is applied to the device in each pixel in the synaptic-like bionic photodetector array to convert the spatial illumination information corresponding to each pixel into a continuously decaying photocurrent signal, and multiple frames of illumination information of the dynamic target are continuously received during the continuous decay of the photocurrent, and after the last frame of illumination information is input, the photocurrent signal of each pixel is read to determine a feature map that memorizes information of all historical moments; The brain-like memristor array is used to perform deep neural network operations on the feature map based on Ohm's law and Kirchhoff's law, output a current corresponding to the task result label, and identify dynamic targets based on the current; the current is used to characterize the probability of the task result label.

2. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 1 is characterized in that: The synapse-like bionic photoelectric detector array specifically includes: a plurality of pixels distributed in an array, each pixel including a retina-like photoelectric detector unit.

3. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 2 is characterized in that: Each of the retinal-like photodetector units includes a substrate, a light absorption layer and two crossed finger electrodes stacked in sequence from bottom to top; wherein the light absorption layer area corresponding to the channel formed between the two crossed finger electrodes is the light response area; the light response area is used to provide linear light response and nonlinear attenuation of photocurrent, and regulate the voltage of the electrodes at both ends of each of the retinal-like photodetector units to encode the light response of the retinal-like photodetector unit.

4. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 3 is characterized in that: The substrate is an insulating Al2O3 substrate; the light absorption layer is a MoS2 two-dimensional material; Performing image design on the MoS2 two-dimensional material by photolithography and etching process, and depositing metal electrodes on the surface of the MoS2 two-dimensional material by electron beam evaporation; The metal electrode is the interdigitated electrode.

5. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 3 is characterized in that: The interdigitated electrodes are rectangular cross-finger structures; the two interdigitated electrodes include a positive electrode and a negative electrode; the positive electrode and the negative electrode are arranged in a staggered manner.

6. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 2 is characterized in that: The photoresponsivity of the retina-like photodetector unit is a random fixed value; The photocurrent signal is the product of the photoresponse rate and the light intensity.

7. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 1 is characterized in that: The photocurrent signal exhibits nonlinear continuous attenuation between frames.

8. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 1 is characterized in that: The brain-like memristor array utilizes word line circuits, source line circuits and bit line circuits to form a mesh layout; a memristor unit is arranged at each intersection in the mesh layout; The memristor unit includes a transistor and a memristor; The word line circuit is connected to the gate of the transistor, the source line circuit is connected to the top electrode of the memristor, and the bit line circuit is connected to the bit of the memristor; The bottom electrode of the memristor is also directly connected to the drain of the transistor; the conductive state of the brain-like memristor array is controlled by adjusting the gate voltage of the transistor; the conductive state includes an on state and an off state.

9. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 8 is characterized in that: The top layer of the memristor array is a top electrode layer, the middle layer is a resistive switching layer, and the bottom layer is a bottom electrode layer; The top electrode layer is input with a voltage through row selection, and the bottom electrode layer is output with a current through column selection; The memristor array follows Ohm's law and Kirchhoff's law to complete multiplication and accumulation operations, reads current from each column, completes vector-matrix multiplication operations in the deep neural network, and realizes a hardware deep neural network with integrated storage and computing.

10. The bionic target sensing, storage and computing integrated photoelectric detection system based on continuous photoconductivity according to claim 1, characterized in that: Also includes: Driving circuit; The driving circuit is directly connected to each electrode of the synaptic-type bionic photodetector array to read the photocurrent signal; The driving circuit is directly connected to the top electrode of the brain-like memristor array, and is used to convert the photocurrent signal into a voltage signal, and input the voltage signal into the brain-like memristor array.

Images