Model reasoning accelerator and intelligent visual perception terminal based on FPGA

Abstract

The present invention provides a model reasoning accelerator and an intelligent visual perception terminal based on FPGA, and the model reasoning accelerator based on FPGA includes a convolution operation unit, a data conversion unit, a pooling operation unit, a translation operation unit, a random reorganization operation unit and a fully connected unit. The above technical solution provided by the present invention is based on the idea of ​​synergy between processing hardware and algorithm model, designs the intelligent computing processing unit architecture, and constructs a neural network model suitable for different hardware environments of above-ground and underground scenarios, with a higher level of automation and intelligence.

Description

[0001] This application is a divisional application of the Chinese patent application with application number 202011035175.8, application date September 27, 2020, and application name “Intelligent visual perception terminal, perception method, storage medium and electronic device in underground coal mines”. Technical Field

[0002] The present invention relates to the technical field of intelligent coal mining, and in particular to a model reasoning accelerator and an intelligent visual perception terminal based on FPGA. Background Art

[0003] In order to promote the integrated development of artificial intelligence and the coal industry, improve the level of intelligent unmanned mining in coal mines, and promote high-quality development of the coal industry, various coal mining companies and coal mining equipment suppliers have invested a lot of manpower and material resources in research on intelligent coal mine technology.

[0004] At present, relevant enterprises have conducted preliminary exploration and research on relevant technologies based on intelligent video monitoring of mines, formed a remote visual intervention type intelligent coal mining control system, and promoted the intelligent unmanned mining process of coal mines. In the process of technical implementation, a large amount of underground video image sample data of coal mines is usually collected first, and a certain scale of machine learning training sample database set is organized and established. Then, a data processing center equipped with high-performance GPU hardware devices is built on the ground, and the intelligent perception algorithm model of the underground equipment environment target object is constructed through supervised training based on deep learning and offline iterative optimization. Finally, the constructed robust perception model is deployed on the ground high-performance workstation, and the underground real-time production video monitoring system is connected through a reliable communication network with a certain bandwidth guarantee to perform real-time reasoning perception of the equipment environment status, and the perception results are used for early warning intelligent monitoring of efficient and safe coal production processes. This remote visual intelligent monitoring video system based on the ground data center can, to a certain extent, liberate ground monitoring management personnel from the long-term busy monotonous manual remote monitoring operation process, and improve the intelligent management and control level of remote visual monitoring.

[0005] However, with the advancement of unmanned mining and the improvement of underground intelligent technology, the real-time requirements of intelligent coal mining control systems are becoming increasingly high. The shortcomings of the real-time remote control of the remote visualization intelligent monitoring video system based on the ground data center are becoming increasingly prominent, which will seriously restrict the process of intelligent unmanned mining in the mine. Due to the rigid electrical requirements of the underground intrinsic safety, the high-power hardware equipment of the ground data processing center architecture is difficult to enter the mine for deployment, which makes it difficult to implement the related intelligent perception algorithm models that require high-power computing and processing resources in the mine. At the same time, due to the constraints of the mine scene conditions, one of the necessary conditions for the ground data processing center architecture, a reliable and fast-responding data communication network with sufficient bandwidth, is also difficult to be popularized in major mines in a short time. Therefore, the shortcomings of the remote visualization intelligent monitoring video system based on the ground data center architecture in the automation and intelligent improvement of coal mine production management are becoming increasingly apparent. Summary of the invention

[0006] The present invention aims to provide a model reasoning accelerator and an intelligent visual perception terminal based on FPGA to solve the technical problem of low automation and intelligence level of remote visualization intelligent monitoring video system in the prior art.

[0007] To this end, an embodiment of the present invention provides a model reasoning accelerator implemented based on FPGA, including:

[0008] The convolution operation unit performs matrix-matrix multiplication operations on the monitoring video image and the model grid weight kernel to obtain a high-bitwidth convolution operation feature map;

[0009] A data conversion unit, converting the high-bitwidth convolution operation feature map data to obtain low-bitwidth feature map data;

[0010] A pooling operation unit receives the low-bit width feature map output by the data conversion unit, and obtains a result according to a certain type of function within a pooling kernel range as a pooling output feature map;

[0011] A translation operation unit translates the pooled output feature map along a plurality of directions to directly copy the neighborhood quantity to the center point position to obtain the same number of channel feature maps as the shift direction;

[0012] A random recombination operation unit receives a plurality of channel feature maps output by the translation operation unit, randomly changes the write-back memory addresses thereof, realizes diversified combinations of the channel feature maps, and obtains randomly recombinated channel feature maps;

[0013] The fully connected unit receives the channel feature map output by the random recombination operation unit and maps it to the convolution operation unit to obtain a feature vector corresponding to the perception information.

[0014] In some embodiments of the model inference accelerator implemented based on FPGA, the data conversion unit includes but is not limited to being implemented by a comparator.

[0015] In some embodiments of the FPGA-based model inference accelerator, the comparator performs byte width conversion on each convolution operation result in the convolution operation unit, reduces the byte width of the convolution operation result to obtain a conversion result, and maps the conversion result into a binary tree structure.

[0016] In some embodiments of the FPGA-based model inference accelerator, the pooling operation unit and the translation operation unit are both implemented by a linear cache design.

[0017] In some embodiments of the FPGA-based model inference accelerator, the random recombination operation unit, the random recombination operation of which includes serially merging the data received this time with the data received last time.

[0018] In some embodiments of the present application, an intelligent visual perception terminal is also provided in an underground coal mine, comprising a camera and an intelligent perception computing and processing unit based on visual content, wherein the camera sends the captured monitoring video image to the intelligent perception computing and processing unit for processing, wherein the intelligent perception computing and processing unit comprises the model inference accelerator based on FPGA as described in any of the above items.

[0019] The intelligent visual perception terminal for underground coal mines in some embodiments further includes:

[0020] A wired transmission interface and a wireless transmission interface are used to input the monitoring video image and output the perception result information; wherein the wired transmission interface includes a wired network interface, an HDMI video data output interface, 2 USB interfaces, a Mircro-SD card slot and an SDRAM memory; the wireless transmission interface includes a dual-band wireless network interface and a Bluetooth interface.

[0021] The intelligent visual perception terminal for underground coal mines in some embodiments further includes:

[0022] An intrinsically safe power management module is used to provide power to the camera and the visual content-based intelligent perception computing processing unit. Compared with the prior art, the above technical solution provided by the embodiment of the present invention has at least the following beneficial effects:

[0023] The model reasoning accelerator and intelligent visual perception terminal based on FPGA provided by the embodiments of the present invention are designed and constructed based on the collaborative idea of ​​algorithm model and hardware processing architecture, and are suitable for the hardware environment of different scenarios above and below the mine. An intelligent terminal with edge computing and accelerated reasoning hardware architecture is proposed, which solves the problem of effective deployment and reasoning of deep learning algorithm models on device terminals in scenarios with limited computing and processing resources of unmanned working faces in deep mines, realizes the real-time perception digitization of the spatial position of the environmental state of unmanned working faces in deep mines, and the on-site perception and on-site cognition problems related to equipment target status detection and identification, and provides a reliable solution for the real-time construction of the three-dimensional space of unmanned working faces in deep mines, the underground on-site intelligent monitoring of key equipment target objects, the underground and above-ground two-way intelligent decision-making management, and the construction of an unmanned intelligent production control system for coal mines with automatic and autonomous execution of underground equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a structural block diagram of an FPGA-based model reasoning accelerator according to an embodiment of the present application;

[0025] Figure 2 An intelligent visual perception terminal for underground coal mines provided by an embodiment of the present invention;

[0026] Figure 3 An intelligent visual perception terminal for underground coal mines provided by another embodiment of the present invention;

[0027] Figure 4 A flow chart of an intelligent visual perception method for underground coal mines provided by an embodiment of the present invention;

[0028] Figure 5 A deployment flow chart of an intelligent visual perception algorithm for underground coal mines provided by an embodiment of the present invention;

[0029] Figure 6 A schematic diagram of the hardware connection relationship of an electronic device for executing an intelligent visual perception method in an underground coal mine provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the orientations or positional relationships indicated by the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. are based on the orientations or positional relationships shown in the accompanying drawings, and are only simplified descriptions for the convenience of describing the present invention, rather than indicating or implying that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. Among them, the terms "first position" and "second position" are two different positions.

[0031] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. Unless there are contradictions between the various technical solutions in the following embodiments provided in this application, different technical solutions can be combined with each other, and the technical features therein can be replaced with each other.

[0032] The embodiment of the present invention provides a model reasoning accelerator based on FPGA, such as Figure 1 As shown, including:

[0033] The convolution operation unit performs matrix-matrix multiplication operations on the monitoring video image and the model grid weight kernel to obtain a high-bitwidth convolution operation feature map.

[0034] The data conversion unit converts the high-bitwidth convolution operation feature map data to obtain low-bitwidth feature map data.

[0035] The pooling operation unit receives the low-bit width feature map output by the data conversion unit, and obtains a result according to a certain type of function within the range of the pooling kernel as a pooling output feature map.

[0036] The translation operation unit translates the pooled output feature map along several directions to directly copy the neighborhood quantity to the center point position to obtain the same number of channel feature maps as the shift direction.

[0037] The random recombination operation unit receives a plurality of channel feature maps output by the translation operation unit, randomly changes the write-back memory addresses thereof, realizes diversified combinations of the channel feature maps, and obtains the randomly recombinated channel feature maps.

[0038] The fully connected unit receives the channel feature map output by the random recombination operation unit and maps it to the convolution operation unit to obtain a feature vector corresponding to the perception information.

[0039] This solution is based on the collaborative idea of ​​algorithm model and hardware processing architecture. It designs and builds a neural network model suitable for different hardware environments above and below the surface. It proposes an intelligent terminal with edge computing and accelerated reasoning hardware architecture, which solves the problem of effective deployment and reasoning of deep learning algorithm models on device terminals in scenarios with limited computing and processing resources in unmanned working faces in deep wells.

[0040] Some embodiments of the present invention provide an intelligent visual perception terminal for underground coal mines, such as Figure 2 and Figure 3 As shown, it includes a camera 101 and a visual content-based intelligent perception computing processing unit 100, wherein the camera 101 sends the captured surveillance video image to the intelligent perception computing processing unit 100 for processing, wherein the intelligent perception computing processing unit 100 includes a model reasoning accelerator 102 implemented based on FPGA as in the above-mentioned embodiment, wherein the FPGA-based model reasoning accelerator 102 includes:

[0041] The convolution operation unit performs matrix-matrix multiplication operation on the monitoring video image and the model grid weight kernel to obtain a high-bit width convolution operation feature map.

[0042] A data conversion unit converts the high-bitwidth convolution operation feature map data to obtain low-bitwidth feature map data. The data conversion unit is implemented by a comparator, but is not limited to being implemented by a comparator. The comparator performs byte width conversion on each convolution operation result in the convolution operation unit, reduces the byte width of the convolution operation result to obtain a conversion result, and then maps the conversion result into a binary tree structure. The convolution operation feature map data is converted from a high byte width to a low byte width, which can simplify the calculation process.

[0043] A pooling operation unit, which is designed with a linear cache to achieve fast calculation, receives the low-bit width feature map output by the data conversion unit, and obtains a result according to a certain type of function within the range of the pooling kernel as a pooling output feature map.

[0044] The translation operation unit translates the pooled output feature map in several directions to copy the neighborhood quantity directly to the center point position, and obtains the same number of channel feature maps as the shift direction. This unit is implemented by a linear cache design, and adds a row of zero-filled items in the horizontal direction and a column of zero-filled items in the vertical direction to the visual feature data output by the merging operation unit, thereby obtaining visual feature data with fill items, that is, the input visual feature data will add 1 column of zero-filled items in the width and height directions respectively, thereby obtaining visual features of width+1 and height+1 sizes. The moving direction is determined according to the channel index value, and the moving directions of different channels are different. After each sliding window calculation after the movement, a value is obtained.

[0045] The random reorganization operation unit receives several channel feature maps output by the translation operation unit, randomly changes the write-back memory address, realizes diversified combinations of channel feature maps, and obtains randomly reorganized channel feature maps. Specifically, the data received this time is combined in series with the data received last time as the visual feature data after this series combination; during the neural network calculation process, the processing unit needs to combine the last calculation result with the current calculation result in series to obtain the final result. The module realizes channel adjustment and information exchange by controlling the offset address of the two calculation results in the series combination process. In order to improve efficiency, this processing unit ensures that when calculating the current result, the CPU can simultaneously complete the copy of the last calculation result, and then connect and store them according to the offset address.

[0046] The fully connected unit receives the channel feature map output by the random recombination operation unit and maps it to the convolution operation unit to obtain a feature vector corresponding to the perception information.

[0047] The visual content-based intelligent perception computing processing unit adopts a quad-core 64-bit ARM Cortex-A72 architecture CPU with a main frequency of 1.5 GHz. Figure 3 As shown, the above-mentioned intelligent visual perception terminal for underground coal mines is configured with: a wired transmission interface and a wireless transmission interface for inputting the monitoring video image and outputting the perception result information; wherein the wired transmission interface includes a wired network interface, an HDMI video data output interface, two USB interfaces, a Mircro-SD card slot and an SDRAM memory; the wireless transmission interface includes a dual-band wireless network interface and a Bluetooth interface. Therefore, the terminal in the above scheme is fully equipped with SDRAM memory, has a Mircro-SD card slot for loading the operating system and data storage, supports Gigabit Ethernet, has a 2.4GHz / 5GHz dual-band 802.11ac wireless network, supports Bluetooth 5.0, provides two USB3.0 interfaces, and supports a high-definition HDMI video data output interface.

[0048] Optionally, the above-mentioned coal mine underground intelligent visual perception terminal also includes: an intrinsically safe power management module, which is used to provide power to the camera and the visual content-based intelligent perception computing processing unit. The intrinsically safe power management module ensures that the electrical characteristics of the terminal system meet the rigid requirements of intrinsic safety for underground coal mine work, so that the terminal's electrical energy and thermal energy are low enough not to cause explosive gases to burn. The principle is: by limiting various circuit parameters or taking protective measures to limit the circuit's spark discharge energy and thermal energy, the electric sparks and thermal effects generated in normal operation and specified fault conditions cannot ignite the explosive mixture in the surrounding environment, thereby achieving electrical intrinsic safety.

[0049] Some embodiments of the present invention also provide an intelligent visual perception method for underground coal mines, such as Figure 4 As shown, the following steps are included:

[0050] S101: construct a model network structure space, and use sample data to train the model in the model network structure space, and use the trained model set as a set of candidate models. The sample data includes monitoring video images and annotated data labels in the monitoring video images.

[0051] S102: Selecting a specific model that is compatible with the actual visual perception application hardware environment from the set of candidate models.

[0052] S103: Inputting the actual application monitoring video image into the selected model network, and obtaining a perception result corresponding to the actual application monitoring video image according to the inference result of the selected model network.

[0053] Specifically, Figure 5 The terminal perception algorithm model construction and deployment are shown in Figure 5 As shown in the left half of , K represents the convolution kernel size (Kmax is the maximum value), D represents the network depth (Dmax is the maximum value), and W represents the number of channels (Wmax is the maximum value). In the process of building the terminal perception algorithm model, the input visual image meets different resolution sizes. The specific steps are described as follows:

[0054] ① Define the model network sample space. According to the convolution kernel size set {K1, K2, ..., K max}, for example {3 , 5 ,..., 9}, that is, the size of the convolution kernel is {3×3, 5×5, 7×7, 9×9}; the network depth set {D 1, D 2,..., D max}, for example, {4,6,8}; the set of channel numbers {W1,W2,…,W max}, such as {3,4,6}; and the resolution size set of different visual images {R1,R2,...,R max}, for example, {256,264,…,Resolution}, defines the network sample space, where Resolution represents the source image resolution.

[0055] ② Initial training of the largest-scale main model network. Through training and learning, the largest-scale model network is automatically constructed, whose convolution kernel size is the maximum size Kmax in the convolution kernel set, the network depth is the maximum number of layers Dmax in the depth number set, the number of channels is the maximum value Wmax in the channel number set, and the resolution size of the visual image is the maximum number Rmax in the resolution set.

[0056] ③ Adaptive training and construction of sub-model network collections. Starting from the largest main model network structure, the scale of the network structure is gradually reduced. Based on the principle of avoiding correlation reasoning between network models, the model network structure space should be exhausted to complete the adaptive training and learning of each sub-model network, and realize the one-time training and construction of the model network collection; after the above adaptive iterative machine learning, a model network collection suitable for perception of diversified hardware platforms is finally obtained to adapt to different hardware system deployments. Among them, the weights between the main model network and the sub-model network are shared.

[0057] ④ After the above-mentioned adaptive iterative machine learning, a set of perception sub-model networks suitable for diverse hardware platforms is finally obtained, which can meet the deployment requirements of different hardware systems.

[0058] ⑤ In the model network terminal pre-deployment stage, the specific model is selected from the model network set by the pre-trained prediction guide, and the specific model is imported into the terminal platform where the actual visual intelligent perception application is located.

[0059] After completing the model network set construction and model network terminal pre-deployment preparation, implement the personalized deployment of the visual perception algorithm model on the underground terminal system platform to realize the visual intelligent on-site perception application. Figure 5 As shown in the right part of the figure, the main steps are as follows:

[0060] (1) Initialization of the application, that is, checking and confirming the environment required for the current program to run;

[0061] Load the configuration parameters required for the parser to run, such as batch_size and other optional configuration parameters in the inference process;

[0062] (2) Read the specific model network required by the loading terminal;

[0063] (3) Create an inference engine based on the loaded model network;

[0064] (4) Perform edge reasoning on the terminal system according to the requirements of the inference engine;

[0065] (5) Output the results of visual intelligent perception applications.

[0066] The above scheme, based on the objective insufficiency of the relevant technologies of the remote visualization intelligent monitoring video system based on the existing ground data center architecture concept, and based on the idea of ​​synergy between hardware structure and software algorithm, proposes an intelligent visual perception terminal and system method for underground coal mines. Firstly, based on the FPGA design, a terminal with edge computing and reasoning capabilities is implemented underground in coal mines, which gives the underground coal mine vision on-site intelligent perception capabilities and reduces the transmission delay of data information. Secondly, on the basis of the implementation of the terminal hardware system, a lightweight algorithm model suitable for the underground terminal hardware environment is constructed through adaptive machine learning to solve the hardware performance limitation of the underground algorithm model deployment, realize the one-time construction of the algorithm model, and meet the needs of diversified hardware platform deployment and reasoning applications.

[0067] The above proposed terminal perception algorithm model construction, deployment, and perception application method system. Through adaptive machine learning, a lightweight algorithm model suitable for the underground terminal hardware environment is constructed to solve the hardware performance limitation of the underground algorithm model deployment, realize the one-time construction of the algorithm model, and meet the deployment of diversified hardware platforms. In the deployment and reasoning application stage, the system selects a specific model network for deployment reasoning and perception application according to the hardware platform and conditional constraint matching.

[0068] Some embodiments of the present invention provide a storage medium, which can be a computer-usable storage medium (including but not limited to disk storage, CD-ROM, optical storage, etc.), in which program instructions are stored. After the computer reads the program instructions, it executes the intelligent visual perception method for underground coal mines described in any of the above technical solutions.

[0069] Figure 6 2 is a schematic diagram of the hardware structure of an electronic device for executing an intelligent visual perception method for underground coal mines provided in this embodiment. The device includes: one or more processors 201 and a memory 202. Figure 6 A processor 201 is taken as an example. The device for executing the intelligent visual perception method in coal mines may also include: an input device 203 and an output device 204. The processor 201, the memory 202, the input device 203 and the output device 204 may be connected via a bus or other means. Figure 5In the example, the bus connection is used. The memory 202 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 201 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 202, that is, the intelligent visual perception method for underground coal mines of the above method embodiment is realized.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent visual perception terminal for underground coal mines, characterized in that: The system comprises a camera and a visual content-based intelligent perception computing and processing unit, wherein the camera sends the captured surveillance video image to the intelligent perception computing and processing unit for processing, wherein the intelligent perception computing and processing unit comprises a model reasoning accelerator implemented based on FPGA; the accelerator comprises: The convolution operation unit performs matrix-matrix multiplication operations on the monitoring video image and the model grid weight kernel to obtain a high-bitwidth convolution operation feature map; A data conversion unit, converting the high-bitwidth convolution operation feature map data to obtain low-bitwidth feature map data; A pooling operation unit receives the low-bit width feature map output by the data conversion unit, and obtains a result according to a certain type of function within a pooling kernel range as a pooling output feature map; A translation operation unit translates the pooled output feature map along a plurality of directions to directly copy the neighborhood quantity to the center point position to obtain the same number of channel feature maps as the shift direction; A random recombination operation unit receives a plurality of channel feature maps output by the translation operation unit, randomly changes the write-back memory address thereof, realizes a diversified combination of the channel feature maps, and obtains a randomly recombinated channel feature map; the random recombination includes serially merging the data received this time with the data received last time; A fully connected unit receives the channel feature map output by the random recombination operation unit and maps it to the convolution operation unit to obtain a feature vector corresponding to the perception information; The coal mine underground intelligent visual perception terminal imports a specific model, and the specific model is obtained in the following way: Define the model network structure sample space: define the network sample space according to the set of convolution kernel sizes, network depths, number of channels, and resolution sizes of different visual images of the network model structure; Initial training of the largest main model network: Automatically build the largest model network through training learning, whose convolution kernel size is the largest size in the convolution kernel size set, network depth is the maximum number of layers in the depth set, number of channels is the maximum value in the channel number set, and resolution size of the visual image is the maximum number in the resolution set; Adaptive training and construction of sub-model network collections: Starting from the largest main model network structure, the network structure scale is gradually reduced. Based on the principle of avoiding correlation reasoning between network models, the model network structure space should be exhausted to complete the adaptive training and learning of each sub-model network, so as to achieve one-time training and construction of the model network collection. After the above adaptive iterative machine learning, a model network collection suitable for perception of diverse hardware platforms is finally obtained to adapt to different hardware system deployments. During the model network terminal pre-deployment phase, the specific model is selected by optimizing the matching from a set of model networks suitable for perception of diverse hardware platforms through a pre-trained prediction director.

2. The intelligent visual perception terminal for underground coal mines according to claim 1, characterized in that: Also includes: A wired transmission interface and a wireless transmission interface are used to input the monitoring video image and output the perception result information; wherein the wired transmission interface includes a wired network interface, an HDMI video data output interface, 2 USB interfaces, a Mircro-SD card slot and an SDRAM memory; the wireless transmission interface includes a dual-band wireless network interface and a Bluetooth interface.

3. The intelligent visual perception terminal for underground coal mines according to claim 2, characterized in that: Also includes: An intrinsically safe power management module, wherein the intrinsically safe power management module is used to provide electrical energy for the camera and the visual content-based intelligent perception computing processing unit.

4. The intelligent visual perception terminal for underground coal mines according to any one of claims 1 to 3, characterized in that: In the accelerator: The data conversion unit includes but is not limited to being implemented by a comparator.

5. The intelligent visual perception terminal for underground coal mines according to claim 4 is characterized in that: The comparator performs byte width conversion on each convolution operation result in the convolution operation unit, reduces the byte width of the convolution operation result to obtain a conversion result, and maps the conversion result into a binary tree structure.

6. The intelligent visual perception terminal for underground coal mines according to claim 5, characterized in that: The pooling operation unit and the translation operation unit are both implemented by linear cache design.

Images