Nonlinear convolution vision acquisition method and system

Abstract

The invention belongs to the field of semiconductor photoelectric devices and intelligent vision, and particularly discloses a nonlinear convolution vision obtaining method and system. The system comprises a plurality of nonlinear convolution kernels formed by nonlinear units; the nonlinear unit is a processing unit for converting an input signal into an output signal in the form of a nonlinear function; one nonlinear unit serves as a convolution kernel pixel unit; the nonlinear function is an exponential function comprising a coefficient item and an exponential item, or a logarithmic function comprising a coefficient item and a base item; the selection of coefficient terms in the nonlinear unit is used for realizing high-pass or low-pass filtering of the image; the selection of the base number item or the index item is used for realizing the contrast adjustment of the image. According to the method, in the calculation process of a nonlinear convolution kernel, contrast enhancement based on a nonlinear term can be realized so as to complete enhancement of a dark image, and functions of Gaussian denoising, edge extraction and the like can also be realized based on low-pass filtering and high-pass filtering functions of a coefficient term.

Description

Technical Field

[0001] The present application belongs to the field of semiconductor optoelectronic devices and intelligent vision, and more specifically, to a nonlinear convolution vision acquisition method and system. Background Art

[0002] With the popularity of electric vehicles, autonomous driving based on machine vision technology is regarded as an important cornerstone of the development of today's automotive technology. As a typical edge computing paradigm, autonomous driving inevitably processes information in various complex lighting environments, the most important of which is the imaging and recognition problems in extremely dark environments and non-uniform and indirect lighting. Traditional machine vision systems require complex algorithms and huge hardware overhead to deal with these problems. This is because they mainly use two methods to deal with such complex lighting environment recognition problems: 1. When a traditional charge-coupled device (CCD) is used for imaging, its light response is fixed. In order to adapt to the illumination in a dark environment, there are two ways: increasing the sensitivity and extending the exposure time. However, increasing the sensitivity requires the circuit to amplify the signal, which will inevitably introduce noise and reduce the image quality; while extending the exposure time will restrict fast image capture, which is very unfavorable for fast-moving imaging such as autonomous driving, and will cause blurred images.

[0003] Second, when traditional machine vision algorithms process images after imaging, in order to improve the low contrast problem caused by complex lighting, it is often necessary to use nonlinear changes in intensity to process the image. The most typical ones are exponential changes and logarithmic changes. However, this nonlinear transformation has a large computational overhead for digital logic circuits, which is not conducive to edge computing such as autonomous driving. Therefore, it is necessary to move the adjustable light response and nonlinear processing from the back end to the front end, or even directly complete them in the image perception process. At present, in terms of device performance, no device can simultaneously complete the two functions of adjustable light response and nonlinear processing; in terms of calculation methods, there is currently no hardware-friendly calculation method that allows nonlinear devices to exert their visual computing capabilities from more angles. Therefore, it is necessary to carry out collaborative innovation in calculation methods and hardware systems to achieve high-speed, high-compact, and low-energy visual architectures and hardware systems for complex lighting scenes. Summary of the invention

[0004] In view of the defects of the prior art, the purpose of this application is to provide a nonlinear convolution vision acquisition method and system, aiming to solve the problem that there is no device that can simultaneously achieve adjustable light response and nonlinear image processing, resulting in low image resolution in complex lighting environments.

[0005] To achieve the above-mentioned purpose, in a first aspect, the present application provides a nonlinear convolution vision system, comprising a plurality of nonlinear convolution kernels composed of nonlinear units; the nonlinear unit is a processing unit that converts an input signal into an output signal in the form of a nonlinear function; a nonlinear unit serves as a convolution kernel pixel unit; the nonlinear function is an exponential function including a coefficient term and an exponential term, or a logarithmic function including a coefficient term and a base term; the nonlinear convolution kernel is used to perform nonlinear convolution processing on an image; the selection of the coefficient term in the nonlinear unit is used to achieve high-pass or low-pass filtering of the image; the selection of the base term or the exponential term is used to achieve contrast adjustment of the image; Among them, the nonlinear convolution kernels are arranged in a matrix form, and the length of the two dimensions of the matrix does not exceed the length of the two dimensions of the pixel matrix corresponding to the input image.

[0006] Further preferably, when the coefficient terms in the nonlinear convolution kernel increase or decrease the highest weight of the kernel center based on the second-order differential operator kernel, and the nonlinear function is sublinear or superlinear, the nonlinear convolution kernel is used to achieve sharpening for light blindness or dark blindness.

[0007] Further preferably, a nonlinear unit in a nonlinear convolution kernel includes two reconfigurable nonlinear photovoltaic heterojunction devices, and the sum of the output currents of the nonlinear units serves as the output of the nonlinear convolution kernel.

[0008] Further preferably, the reconfigurable nonlinear photovoltaic heterojunction device is n-Si / p-AgO x Heterojunction; where 0.4<x<1.2; the thickness of the n-Si layer is 10nm~500um; p-AgO x The layer thickness is 10nm~100nm.

[0009] Further preferably, the p-AgO of the reconfigurable nonlinear photovoltaic heterojunction device of the nonlinear unit x The n-Si direction outputs positive current, the n-Si direction outputs negative current, two reconfigurable nonlinear photovoltaic heterojunction devices are connected in reverse parallel, and the p-AgO x The sum of the positive current output in the direction and the negative current output in the n-Si direction is the output of the nonlinear unit with positive and negative adjustment. The sum of the outputs of all nonlinear units is the output of the nonlinear convolution kernel.

[0010] In a second aspect, the present application provides a nonlinear convolution vision acquisition method, which specifically includes the following steps: Determine the size of the nonlinear convolution kernel, adjust the coefficient term in the nonlinear convolution kernel to achieve high-pass filtering or low-pass filtering of the image; adjust the exponential term or base term in the nonlinear convolution kernel according to the preset contrast of the image; All elements of the image matrix are input into the nonlinear convolution kernel in turn according to the size of the nonlinear convolution kernel for nonlinear mapping. The output results of the mapping of each nonlinear unit on the nonlinear convolution kernel are added to obtain the element output of the image after one convolution. Similarly, the image elements are convolved in turn and the output elements are arranged to obtain the image after nonlinear convolution processing.

[0011] Further preferably, the coefficient term of the nonlinear convolution kernel is increased or decreased based on the second-order differential operator kernel with respect to the highest weight of the kernel center, and the nonlinear function is made sublinear or superlinear, so as to realize the nonlinear convolution kernel to perform light-blind or dark-blind sharpening on the image.

[0012] Further preferably, the coefficient term and exponential term adjustment method in the nonlinear convolution kernel is: Based on the characteristics that the larger the negative voltage and the longer the duration, the higher the degree of ionization, the coefficient term in the photoelectric exponential response function of the reconfigurable nonlinear photovoltaic heterojunction device increases and the exponential term decreases, in the p-AgO reconfigurable nonlinear photovoltaic heterojunction device x A negative voltage is applied to ionize the trap center; the reconfigurable nonlinear photovoltaic heterojunction device is n-Si / p-AgO x Heterojunction; Based on the characteristics that the larger the forward voltage and the longer the duration, the higher the degree of deionization, the coefficient term in the photoelectric exponential response function of the reconfigurable nonlinear photovoltaic heterojunction device decreases and the exponential term increases, in the p-AgO reconfigurable nonlinear photovoltaic heterojunction device x Applying a forward voltage causes deionization at the center of the trap.

[0013] In general, the above technical solutions conceived by this application have the following beneficial effects compared with the prior art: The present application provides a nonlinear convolution vision acquisition method, which utilizes the linear coefficient term and nonlinear term (exponential term or base term) in the nonlinear unit for calculation. In the calculation process of a nonlinear convolution kernel, it can realize contrast enhancement based on nonlinear terms to enhance dark images, and realize low-pass filtering and high-pass filtering based on coefficient terms to complete Gaussian denoising and edge extraction. In particular, it can perform bright blindness and dark blindness sharpening, which is a function that traditional linear kernels cannot perform. Nonlinearity is a non-ideal characteristic of a device, and it has not had much application value before. The nonlinear convolution vision acquisition method provided by the present application is suitable for most nonlinear optoelectronic devices. It is a hardware-friendly convolution vision algorithm that broadens the use scenarios of non-ideal devices.

[0014] The present application provides a nonlinear convolution vision system, which utilizes a reconfigurable nonlinear photovoltaic heterojunction n-Si / p-AgO xThe photoelectric response of the device is exponentially adjustable, and a nonlinear photovoltaic cell regulated by traps is used as a photoelectric calculation unit; at the same time, the device p-AgO x The positive current output from the port is added to the negative current output from the n-Si port of another device to obtain a positive and negative adjustable nonlinear convolution kernel unit output, which is finally expanded to obtain a nonlinear convolution vision system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of a nonlinear convolution vision system provided in an embodiment of the present application; FIG. 2 (a) is a diagram of the optoelectronic performance of a nonlinear convolutional visual computing device unit provided in an embodiment of the present application; FIG2( b ) is a double logarithmic coordinate fitting diagram of the photoelectric response function of the nonlinear photovoltaic heterojunction device provided in the embodiment of the present application in different photoelectric response states, and the fitting data is from FIG2( a ); FIG2 (c) is a parameter of the photoelectric response index function of the nonlinear photovoltaic heterojunction device provided in the embodiment of the present application in different photoelectric response states; wherein the parameter is derived from the fitting result of FIG2 (b), and is affected by the space charge limited photocurrent. When the nonlinear term is 1, it indicates that there is no space charge influence, and when it is less than 1, it indicates that there is influence, and the influence increases as it decreases; Figure 3 This is a comparison diagram of the algorithms and effects of the nonlinear convolution algorithm based on high-pass filtering and the traditional convolution algorithm in image processing provided by the embodiment of the present application; Figure 4 This is a comparison chart of the algorithms and effects of the nonlinear convolution algorithm based on low-pass filtering and the general vision algorithm in image processing provided by the embodiment of the present application; Figure 5 This is a comparison chart of the classification effects of the nonlinear convolution algorithm for edge extraction provided in the embodiment of the present application and the traditional convolution algorithm on a public data set. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0017] The term "and / or" in this article is a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The symbol " / " in this article indicates that the associated objects are in an or relationship, for example, A / B means A or B.

[0018] The terms “first”, “second” and the like in the specification and claims herein are used to distinguish different objects rather than to describe a specific order of the objects.

[0019] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0020] In the description of the embodiments of the present application, unless otherwise specified, “plurality” means two or more than two.

[0021] The present application provides a nonlinear convolution visual computing method and system, and proposes a nonlinear convolution kernel, each unit of which has an adjustable nonlinear functional relationship between input and output signals, and this function is an exponential function containing exponential terms and coefficient terms; the selection of exponential terms can enhance the contrast of image intensity, and the selection of coefficient terms can complete the high-pass or low-pass filtering function, thereby realizing multi-dimensional image processing on one kernel. Specifically, in the hardware system for implementing this method, a reconfigurable nonlinear photovoltaic heterojunction device is used as a convolution kernel pixel unit, and the device uses an electrical signal to control the change in the degree of trap ionization to control the exponential function of input light power-output photocurrent; the system based on this method has strong robustness in processing complex illumination images, and can complete processing tasks such as edge extraction, Gaussian denoising and bright-blind sharpening under various conditions.

[0022] More specifically, Figure 1 As shown, the present application provides a nonlinear convolution vision system, including a nonlinear kernel composed of nonlinear units, and the nonlinear kernel performs nonlinear convolution processing on the image; the nonlinear unit converts the input signal into an output signal in the form of a nonlinear function, and the nonlinear function can be a coefficient term and exponential terms Exponential function , or including coefficient terms and base terms The logarithmic function of One of Figure 1 The nonlinear convolution method with exponential function as kernel is shown in; More specifically, the nonlinear kernel composed of the nonlinear units processes the input image signal by means of convolution. According to the selection of the values ​​of different parameter items, the output result will have the high-pass or low-pass filtering effect caused by the multiplication and addition of coefficient items, and the exponential change or logarithmic change effect of the image intensity caused by the exponential item or the base item; Further preferably, the nonlinear convolution kernels are arranged in a matrix form, and the length of two dimensions of the matrix cannot exceed the length of two dimensions of the input image matrix at most, and the minimum length is one; Further preferably, when the coefficient term of the nonlinear convolution kernel increases or decreases a minimum weight value to the highest weight of the kernel center based on the second-order differential operator kernel, and the nonlinear function is made sublinear or superlinear, the nonlinear convolution kernel can perform a bright-blind or dark-blind sharpening effect; Further preferably, the nonlinear convolution processing process is: all elements of the input image matrix are sequentially input into the nonlinear convolution kernel according to the kernel size for nonlinear mapping, the mapping output results of each unit on the kernel are added to obtain the element output of the image after one convolution, and the output elements are sequentially processed and arranged to obtain the final image after nonlinear convolution processing; Example 1 The present application provides a nonlinear convolution vision system, comprising a circuit composed of a reconfigurable nonlinear photovoltaic heterojunction device, wherein the sum of the circuit output currents is used as the output of the convolution kernel; Further preferably, the reconfigurable nonlinear photovoltaic heterojunction device is n-Si / p-AgO x Heterojunction, where 0.4<x<1.2; n-Si layer thickness is 10nm-500um, p-AgO x The layer thickness is 10nm-100nm; Further preferably, the reconfigurable nonlinear photovoltaic heterojunction device p-AgO x The n-Si direction outputs positive current, and the n-Si direction outputs negative current. The sum of the two is a positively and negatively adjustable nonlinear convolution kernel unit output. By expanding multiple such units into an array, a hardware system of a single nonlinear convolution kernel can be obtained.

[0023] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application.

[0024] Example 1 The present application discloses a reconfigurable nonlinear photovoltaic heterojunction device n-Si / p-AgO x , its polymorphic reconfiguration can be achieved through the stimulation of electrical signals; when the heterojunction is reverse biased and the voltage is greater than 0.5V (p-AgO x When a negative voltage is applied, the trap center will be ionized. The greater the voltage and the longer the duration, the higher the degree of ionization. This increased degree of ionization will increase the coefficient term and decrease the exponential term in the photoelectric exponential response function of the device. When the heterojunction is forward biased and the voltage is greater than 0.5V (p-AgO xWhen a forward voltage is applied, deionization will occur at the trap center. The greater the voltage and the longer the duration, the higher the degree of deionization. The increase in the degree of deionization will cause the coefficient term in the photoelectric exponential response function of the device to decrease and the exponential term to increase. In this way, the nonlinear response of the device based on the exponential function is reconfigurable, as shown in Figure 2 (a). At the same time, the response of this exponential function is fitted with double logarithmic coordinates, as shown in Figure 2 (b). Finally, the coefficient terms under different nonlinear states are obtained. With exponential terms , as shown in Figure 2 (c), according to the physical mechanism of space charge limited photocurrent, when the nonlinear term is 1, it means there is no space charge influence, and when it is less than 1, it means there is influence, and as it decreases, the influence increases, and the influence of space charge is caused by trap ionization, which further proves that the working mechanism of the device is closely related to trap ionization; More specifically, the present application provides a nonlinear convolution kernel, each unit of which has an adjustable nonlinear function relationship between input and output signals, and this function is an exponential function containing exponential terms and coefficient terms; the selection of exponential terms can achieve contrast enhancement on image intensity, and the selection of coefficient terms can achieve high-pass or low-pass filtering functions, thereby realizing multi-dimensional image processing on one kernel. Specifically, in the hardware system for implementing this method, a reconfigurable nonlinear photovoltaic heterojunction device is used as a convolution kernel pixel unit, and the device uses electrical signals to regulate the change in the degree of trap ionization to control the exponential function relationship between input optical power and output photocurrent; the system based on this method has strong robustness in processing complex illumination images, and can complete processing tasks such as edge extraction, Gaussian denoising, and bright-blind sharpening under various conditions.

[0025] Example 2 According to the reconfigurable nonlinear photovoltaic heterojunction device proposed in the embodiment, the embodiment of the present application builds the following Figure 1 The hardware system shown in the figure simulates the image high-pass filtering process; The most classic representatives of high-pass filtering are Sobel and sharpening kernel convolution; Figure 3 This is a comparison chart of the nonlinear convolution algorithm and the traditional convolution algorithm in this application for processing images in complex lighting environments with high-pass filtering as the goal; among them, the traditional convolution algorithm is based on the Sobel kernel and the sharpening kernel; Figure 3 The input image is a typical complex lighting scene; Figure 3 Nonlinear convolution kernels and linear convolution kernels are provided; wherein the coefficient terms of the nonlinear edge horizontal and vertical kernels are set according to the linear Sobel horizontal and vertical kernels, which give the same edge extraction function; the nonlinear term is the nonlinear parameter corresponding to the coefficient kernel, which comes from the coefficient term-exponential term corresponding result of each memory state in Figure 2 (c); In the processing kernel for sharpening, based on the setting of the linear sharpening kernel, the present application provides a nonlinear light-blind sharpening kernel; different from the linear sharpening kernel that increases the weight of the kernel center on the basis of the second-order difference to enhance the contrast amplification in bright conditions, the present application reduces the weight of the kernel center on the basis of the second-order difference, and uses nonlinear terms to enhance the contrast of the dark part, so that the bright part is directly reduced to 0, while the dark part will greatly enhance its response, ultimately forming a light-blind (or dark-selective) sharpening effect.

[0026] Figure 3 The convolution processing results corresponding to the nonlinear kernel and the linear kernel are provided; for the high-pass filtering of edge extraction, the nonlinear convolution can stably identify the portrait in the dark environment behind the door, but the linear convolution has difficulty in performing this function; for the high-pass filtering of image sharpening, the nonlinear convolution demonstrates the function of bright-blind sharpening, all bright parts are reduced to 0, but the contrast of the dark parts is significantly amplified and sharpened, while the linear sharpening kernel can only monotonically increase the sharpness of the bright parts; this special function unique to nonlinear convolution will be of great help for dark-selective image processing in some special extreme environments.

[0027] Example 3 According to the reconfigurable nonlinear photovoltaic heterojunction device provided in the embodiment, the embodiment of the present application builds the following Figure 1 The hardware system shown in , and the image low-pass filtering process is simulated; The Gaussian kernel is the most classic representative of low-pass filtering. Figure 4 The comparison chart of the nonlinear convolution algorithm of this application and the traditional convolution algorithm in the processing of complex lighting environment imaging for low-pass filtering is shown, which is mainly based on Gaussian kernel; Figure 4 It provides nonlinear kernels and traditional low-pass filtering algorithms for convolution processing; the traditional low-pass filtering algorithm will perform exponential changes before processing the image, thereby improving the overall contrast of the image; this will cause the image processing to be performed in two steps, increasing the time for data transmission and storage. Different from the general low-pass filtering algorithm, the nonlinear Gaussian kernel has nonlinear characteristics of the device itself, so the nonlinear process is completed synchronously in the convolution and does not need to be performed in steps; at the same time, the most important point is that since the nonlinear kernel has multiple different exponential terms (from 0.2 to 0.74), and the general low-pass filtering algorithm has only one exponential term (0.2), it often causes the brightness to increase too much during the nonlinear change process, which in turn reduces the contrast, and this situation does not often occur with nonlinear kernels with multiple different exponential terms; Figure 4A comparison chart of the input image, the image after nonlinear convolution processing, and the image after general low-pass filtering algorithm processing is provided. It can be seen that the contrast of the input image is 0.0393, the contrast of the image after nonlinear convolution is 0.0475, and the contrast of the image after general low-pass filtering algorithm processing is 0.0337. The contrast of the image after nonlinear convolution processing is significantly better than that of the general low-pass filtering algorithm processing; this is because the nonlinear transformation of the general low-pass filtering algorithm is too monotonous and cannot balance the global contrast in a larger range.

[0028] Example 4 According to the reconfigurable nonlinear photovoltaic heterojunction device provided in the embodiment, the present application constructs the following Figure 1 The hardware system shown in the figure is used to simulate the classification effect of a large-scale public dataset. In order to compare the difference between nonlinear convolution and traditional linear convolution methods in a wider range, the embodiment of the present application selects the ExtYaleBCroppedPNG public dataset as the classification task. This dataset is a classic and widely used complex illumination imaging classification image; the images in the dataset come from illumination at different angles. This angle is the angle between the observation direction and the light input direction, and is represented by azimuth (abbreviated as A) and elevation (abbreviated as E), such as Figure 5 As shown; image nonlinear convolution processing and traditional linear convolution processing methods come from Figure 3 The edge image is obtained by combining the horizontal and vertical results of the edge extraction kernel convolution and outputting them. Specifically, it is obtained by calculating the Euclidean norm of the two, which is a common calculation method in digital image processing.

[0029] Figure 5 The original input image and the images processed by the nonlinear and linear convolution algorithms are provided; the images are arranged from small to large according to the changes in azimuth angle A and elevation angle E, which reflects the increase in the complexity of the image illumination, which can be seen from the input original image; it can be seen from the image processing results that with the continuous increase in the illumination angle, the results after nonlinear convolution processing maintain a very good edge recognition effect, but the results after traditional linear convolution processing are difficult to extract effective information in the dark area; further, the entire data set is arranged from small to large according to the norm of azimuth angle A and elevation angle E, and is evenly divided into 4 sub-data sets (such as Figure 5 As shown in the figure, the nonlinear and linear convolution results of the four sub-datasets are input into a 3-layer neural network after downsampling to classify 30 types of faces (i.e., 30 labels) (as shown in the figure). Figure 5 shown); Figure 5The final classification accuracy is provided. It can be seen that with the increasing complexity of illumination (from sub-datasets 1 to 4), the accuracy of the nonlinear convolution results remains stable at above 90%, while the accuracy of the traditional linear convolution results drops sharply and ultimately fails to effectively output the correct results.

[0030] In summary, compared with the prior art, the present application has the following advantages: The present application provides a nonlinear convolution vision calculation method, which utilizes linear coefficient terms and nonlinear terms (exponential terms or base terms) for calculation. In the calculation process of a kernel, it can realize contrast enhancement based on nonlinear terms to enhance dark images, and realize low-pass filtering and high-pass filtering based on coefficient terms to complete Gaussian denoising and edge extraction. In particular, it can perform bright-blind and dark-blind sharpening, which is a function that traditional linear kernels cannot perform. Nonlinearity is a non-ideal characteristic of a device, which has not had much application value before. This calculation method is applicable to most nonlinear optoelectronic devices. It is a hardware-friendly convolution vision algorithm that broadens the use scenarios of non-ideal devices.

[0031] This application provides a nonlinear convolution vision hardware system that utilizes a reconfigurable nonlinear photovoltaic heterojunction n-Si / p-AgO x The photoelectric response of the device is exponentially adjustable, and a nonlinear photovoltaic cell regulated by traps is used as a photoelectric calculation unit; at the same time, the device p-AgO x The positive current output from the port is added to the negative current output from the n-Si port of another device to obtain a positive and negative adjustable convolution kernel unit output, which is finally expanded to obtain a hardware system of a nonlinear convolution kernel.

[0032] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A nonlinear convolutional vision system, characterized in that: It includes a number of nonlinear convolution kernels composed of nonlinear units; the nonlinear unit is a processing unit that converts an input signal into an output signal in the form of a nonlinear function; a nonlinear unit serves as a convolution kernel pixel unit; the nonlinear function is an exponential function including a coefficient term and an exponential term, or a logarithmic function including a coefficient term and a base term; the nonlinear convolution kernel is used to perform nonlinear convolution processing on an image; the selection of the coefficient term in the nonlinear unit is used to achieve high-pass or low-pass filtering of the image; the selection of the base term or the exponential term is used to achieve contrast adjustment of the image; Among them, the nonlinear convolution kernels are arranged in a matrix form, and the length of the two dimensions of the matrix does not exceed the length of the two dimensions of the pixel matrix corresponding to the input image.

2. The nonlinear convolutional vision system according to claim 1, characterized in that: When the coefficient terms in the nonlinear convolution kernel increase or decrease the highest weight of the kernel center based on the second-order differential operator kernel, and the nonlinear function is sublinear or superlinear, the nonlinear convolution kernel is used to achieve bright-blind or dark-blind sharpening.

3. The nonlinear convolutional vision system according to claim 1, characterized in that: A nonlinear unit in a nonlinear convolution kernel includes two reconfigurable nonlinear photovoltaic heterojunction devices, and the sum of currents output by the nonlinear unit serves as the output of the nonlinear convolution kernel.

4. The nonlinear convolutional vision system according to claim 3, characterized in that: Reconfigurable nonlinear photovoltaic heterojunction device for n-Si / p-AgO x Heterojunction; where 0.4<x<1.2; the thickness of the n-Si layer is 10nm~500um; p-AgO x The layer thickness is 10nm~100nm.

5. The nonlinear convolutional vision system according to claim 4, characterized in that: Reconfigurable nonlinear photovoltaic heterojunction devices based on nonlinear units in p-AgO x The n-Si direction outputs positive current, the n-Si direction outputs negative current, two reconfigurable nonlinear photovoltaic heterojunction devices are connected in reverse parallel, and the p-AgO x The sum of the positive current output in the direction and the negative current output in the n-Si direction is the output of the nonlinear unit with positive and negative adjustment. The sum of the outputs of all nonlinear units is the output of the nonlinear convolution kernel.

6. A nonlinear convolution vision acquisition method based on the nonlinear convolution vision system according to any one of claims 1 to 5, characterized in that: The specific steps include: Determine the size of the nonlinear convolution kernel, adjust the coefficient term in the nonlinear convolution kernel to achieve high-pass filtering or low-pass filtering of the image; adjust the exponential term or base term in the nonlinear convolution kernel according to the preset contrast of the image; All elements of the image matrix are input into the nonlinear convolution kernel in turn according to the size of the nonlinear convolution kernel for nonlinear mapping. The output results of the mapping of each nonlinear unit on the nonlinear convolution kernel are added to obtain the element output of the image after one convolution. Similarly, the image elements are convolved in turn and the output elements are arranged to obtain the image after nonlinear convolution processing.

7. The nonlinear convolution vision acquisition method according to claim 6, characterized in that: The coefficient term of the nonlinear convolution kernel is increased or decreased on the basis of the second-order differential operator kernel for the highest weight of the kernel center, and the nonlinear function is made sublinear or superlinear, so that the nonlinear convolution kernel can perform bright-blind or dark-blind sharpening on the image.

8. The nonlinear convolution vision acquisition method according to claim 6 or 7, characterized in that: The adjustment method of coefficient term and exponential term in nonlinear convolution kernel is: Based on the characteristics that the larger the negative voltage and the longer the duration, the higher the degree of ionization, the coefficient term in the photoelectric exponential response function of the reconfigurable nonlinear photovoltaic heterojunction device increases and the exponential term decreases, in the p-AgO reconfigurable nonlinear photovoltaic heterojunction device x A negative voltage is applied to ionize the trap center; the reconfigurable nonlinear photovoltaic heterojunction device is n-Si / p-AgO x Heterojunction; Based on the characteristics that the larger the forward voltage and the longer the duration, the higher the degree of deionization, the coefficient term in the photoelectric exponential response function of the reconfigurable nonlinear photovoltaic heterojunction device decreases and the exponential term increases, in the p-AgO reconfigurable nonlinear photovoltaic heterojunction device x Applying a forward voltage causes deionization at the center of the trap.

Images