A computer image processing system using intelligent recognition technology

Abstract

The application discloses a computer image processing system adopting intelligent identification technology, which comprises an image acquisition module, a preprocessing module and an intelligent identification module; the image acquisition module comprises a camera, a high-resolution sensor, an infrared / thermal imaging device and a 3D laser radar, and realizes low-delay data stream transmission through a 5G / optical fiber network; the preprocessing module is used for dynamic light correction and noise suppression based on an adaptive filter; the intelligent identification module comprises a feature extraction unit; the application relates to the technical field of image processing devices, in particular to a computer image processing system adopting intelligent identification technology, which realizes low-delay data stream transmission through a 5G / optical fiber network, combines various image acquisition devices and can quickly and accurately acquire image data; the system is provided with a supporting part, a scanning assembly and a flattening assembly, can automatically detect unflattened positions of paper and perform flattening treatment, and improves the efficiency and quality of paper image acquisition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing devices, in particular to a computer image processing system adopting intelligent recognition technology. BACKGROUND

[0002] In today's digital age, computer image processing technology plays a key role in many fields such as industrial detection, intelligent security, medical image analysis, etc. However, the existing computer image processing system has some shortcomings.

[0003] Traditional image acquisition methods often have high data transmission delay, which cannot meet the real-time requirements of high application scenarios. In the image preprocessing stage, light changes and noise interference will seriously affect the accuracy of subsequent processing, and the existing correction and suppression methods have limited effect.

[0004] In terms of intelligent recognition, some deep learning models have high complexity and large amount of calculation, which requires harsh hardware resources and is not conducive to deployment on resource-constrained devices. The semantic understanding ability is insufficient, and it is difficult to accurately generate image descriptions that meet the actual situation. The abnormal detection method is not flexible enough to effectively identify outliers in complex scenarios. At the same time, misidentification is common, and there is a lack of effective correction mechanism.

[0005] In addition, for paper image processing, there is a lack of system that can automatically flatten the paper and accurately scan it, and manual flattening is inefficient and prone to errors. Therefore, for paper flattening direction, a computer image processing system adopting intelligent recognition technology is designed. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a computer image processing system adopting intelligent recognition technology, which realizes low-delay data stream transmission through 5G / optical fiber network, combines multiple image acquisition devices, and can quickly and accurately acquire image data; The system is equipped with a support part, a scanning assembly and a flattening assembly, which can automatically detect the unflattened part of the paper and perform flattening processing, thereby improving the efficiency and quality of paper image acquisition.

[0007] A computer image processing system adopting intelligent recognition technology, comprising:

[0008] An image acquisition module, the image acquisition module comprising a camera, a high-resolution sensor, an infrared / thermal imaging device, and a 3D laser radar, and realizing low-delay data stream transmission through 5G / optical fiber network;

[0009] A preprocessing module for dynamic light correction and noise suppression based on an adaptive filter;

[0010] The intelligent recognition module includes a feature extraction unit, which uses a lightweight deep learning model combined with an attention mechanism to achieve feature extraction. The lightweight deep learning model is MobileNetV3, and the attention mechanism is ECA-Net or Dynamic Spatial Attention DSAM.

[0011] The semantic understanding unit is used to generate image descriptions;

[0012] Anomaly detection unit identifies outliers based on Autoencoder;

[0013] Decision output unit, used to generate structured reports;

[0014] The post-processing module corrects misidentifications based on contextual information.

[0015] The controller is connected to the image acquisition module, preprocessing module, intelligent recognition module, semantic understanding unit, anomaly detection unit, decision output unit, and post-processing module, respectively. It is used to control and coordinate the data transmission and working sequence between the modules to ensure that the entire computer image processing system operates in a predetermined manner. This includes, but is not limited to, instructing the preprocessing module to perform illumination correction and noise suppression of corresponding parameters based on the data acquired by the image acquisition module; guiding the semantic understanding unit to perform targeted image description generation based on the feature extraction results of the intelligent recognition module; and controlling the decision output unit to decide whether to generate a structured report based on the outlier identification results of the anomaly detection unit.

[0016] As a further limitation of this technical solution, in the feature extraction unit, MobileNetV3 adopts depthwise separable convolution, inverse residual structure, hardware-friendly activation function h-swish and SE module, and optimizes the model structure through channel pruning, layer fusion and adjustment of width scaling factor. During training, QAT is used to simulate 8-bit integer calculation, and during deployment, the weights are quantized with INT8 and the activation values ​​are dynamically range quantized.

[0017] As a further limitation of this technical solution, the semantic understanding unit adopts an encoder-decoder framework or a multimodal pre-trained model architecture, wherein the visual feature extraction model includes one or more of ResNet-101, ViT-L / 16, CLIP-ViT, and Faster R-CNN, and the text generation module adopts one of LSTM+Attention, Transformer, and GPT-2, and the visual and text features are fused by calculating a cross-modal attention weight matrix.

[0018] As a further limitation of this technical solution, the Autoencoder of the anomaly detection unit includes a basic Autoencoder structure and one or more improved variants of VAE, ConvAE, LSTM-AE, and Attention AE. During training, one or more of mean squared error, weighted MSE, adversarial loss, and Mahalanobis distance are used as loss functions. The anomaly judgment threshold is selected by one or more methods of quantile method, K-Sigma, extreme value theory, and dynamic threshold.

[0019] As a further limitation of this technical solution, the post-processing module performs misidentification correction based on Conditional Random Field (CRF). By modeling the contextual dependencies between pixels / regions, it balances univariate potential and binary potential. The univariate potential is based on the initial classification confidence, and the binary potential is based on the similarity constraints of spatial features, color features, and semantic features.

[0020] As a further limitation of this technical solution, it includes a support part, on both sides of which the two ends of the bracket are fixed respectively. The bracket is arched, and the upper lower part of the bracket is provided with a fixed end of the lifting part. The movable end of the lifting part is fixedly connected to a flattening component, which is used for flattening the paper.

[0021] The support is also connected to a scanning component, which is used to monitor any unflattened areas of the paper placed inside the support.

[0022] As a further limitation of this technical solution, the flattening assembly includes a limiting part, a limiting slot, a connecting rod, a fixed plate, a rotating plate, a flattening part, a fixed shaft, and a driving part. The flattening part consists of several sector-shaped blocks, all of which can be enclosed to form a circular plate with a central hole. The fixed plate fixes the movable end of the lifting part. A positioning shaft is fixedly connected to the lower center of the fixed plate. The positioning shaft matches the central hole of the circular plate formed by all the sector-shaped blocks. The outer arc surface of the positioning shaft can abut against the inner arc surface of the sector-shaped blocks. A rotating plate is provided on the lower side of the fixed plate. The rotating plate is circular. The rotating plate and the... The positioning shaft is rotatably connected, and the flattening part is provided on the lower side of the rotating plate. One end of the drive rod is fixedly connected to the middle part of the outer arc surface of the flattening part. The drive rod is bent, and the other end of the drive rod is slidably disposed in the limiting slot. The limiting slot is fixed at the upper edge of the fixed plate. Each drive rod is rotatably connected to one end of the connecting rod, and the other end of each connecting rod is rotatably connected to the fixed shaft. The fixed shaft is circumferentially fixed at the upper edge of the rotating plate. Rollers are provided at the lower end of the positioning shaft and the lower side of each flattening part, and the lower sides of the rollers are flush.

[0023] As a further limitation of this technical solution, one of the drive rods has a fixed protrusion at the other end, the protrusion is rotatably connected to the telescopic rod of the first electric push rod, the outer shell end of the first electric push rod is rotatably connected to the upper side of the fixed disk, and a number of circumferentially evenly arranged fixed shafts are fixedly provided at the upper edge of the fixed disk, the fixed shafts being set at the middle position corresponding to the outward unfolding movement of the flattened part.

[0024] As a further limitation of this technical solution, there is at least one lifting part, which is an electric push rod. The outer shell of the second electric push rod is fixed to the bracket, and the telescopic rod of the second electric push rod is fixed to the center of the rotating plate.

[0025] Fixed blocks are fixedly connected to both ends of one side of the support unit. The fixed blocks are rotatably connected to both ends of the connecting screw. The screw is screwed to a moving block. The moving block is fixedly connected to the center of the scanning unit. The scanning unit is located on the upper side inside the support unit. Several evenly arranged visual sensors are fixedly installed on the lower side of the scanning unit. A motor is fixed to one end of the screw. The output shaft of the motor is fixedly connected to the fixed block. The visual sensor is the camera. The controller transmits data with the visual sensor. The controller is electrically connected to the lifting unit, the drive unit, and the motor.

[0026] Compared with the prior art, the advantages and positive effects of the present invention are:

[0027] (1) Low-latency data stream transmission is achieved by using 5G / fiber optic networks. Combined with various image acquisition devices, image data can be acquired quickly and accurately. Dynamic illumination correction and noise suppression based on adaptive filters effectively improve image quality and provide a good foundation for subsequent processing.

[0028] (2) The lightweight deep learning model combined with the attention mechanism reduces the model complexity and computational load while ensuring recognition accuracy, making it easy to deploy on resource-constrained devices. It adopts a variety of advanced visual feature extraction models and text generation modules, which can accurately generate image descriptions and improve the semantic understanding ability of the system. The outlier recognition based on Autoencoder, combined with a variety of improved variants and loss functions, can flexibly adapt to the anomaly detection needs in different scenarios. The post-processing module corrects misidentification based on Conditional Random Field (CRF) and effectively improves the recognition accuracy by modeling contextual dependencies.

[0029] (3) The system is equipped with a support unit, a scanning component and a flattening component, which can automatically detect the unflattened parts of the paper and flatten them, thus improving the efficiency and quality of paper image acquisition. Attached Figure Description

[0030] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0031] Figure 1 This is a flowchart of the present invention;

[0032] Figure 2 The three-dimensional representation of the present invention Figure One ;

[0033] Figure 3 The three-dimensional representation of the present invention Figure Two ;

[0034] Figure 4 The three-dimensional representation of the present invention Figure Three .

[0035] In the diagram: 1. Bracket; 2. Lifting part; 3. Limiting part; 4. Limiting slot; 5. Connecting rod; 6. Rotating plate; 7. Flattening part; 701. Roller; 8. Fixed shaft; 9. Support part; 10. Fixed block; 11. Screw; 12. Moving block; 13. Motor; 14. Drive part; 15. Scanning part; 16. Fixed plate; 17. Drive rod; 18. Fixed shaft. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] A computer image processing system employing intelligent recognition technology includes:

[0038] The image acquisition module includes a camera, a high-resolution sensor, an infrared / thermal imaging device, and a 3D LiDAR, and achieves low-latency data stream transmission via a 5G / fiber optic network. The camera, high-resolution sensor, infrared / thermal imaging device, 3D LiDAR, and other devices in the image acquisition module acquire image data and transmit the data to the system with low latency via a 5G / fiber optic network.

[0039] The preprocessing module performs dynamic illumination correction and noise suppression based on adaptive filters on the acquired images to improve image quality.

[0040] The intelligent recognition module includes a feature extraction unit, which uses a lightweight deep learning model combined with an attention mechanism to extract features. The lightweight deep learning model is MobileNetV3, and the attention mechanism is ECA-Net or Dynamic Spatial Attention DSAM. The feature extraction unit of the intelligent recognition module uses a lightweight deep learning model (MobileNetV3) combined with an attention mechanism (ECA-Net or Dynamic Spatial Attention DSAM) to extract features.

[0041] The semantic understanding unit is used to generate image descriptions; the semantic understanding unit generates image descriptions based on the feature extraction results.

[0042] The anomaly detection unit identifies outliers based on the Autoencoder; subsequently...

[0043] Decision output unit, used to generate structured reports;

[0044] The post-processing module corrects misidentifications based on contextual information; the processing module corrects misidentifications based on Conditional Random Field (CRF) by balancing univariate and binary potentials by modeling the contextual dependencies between pixels / regions.

[0045] The controller is connected to the image acquisition module, preprocessing module, intelligent recognition module, semantic understanding unit, anomaly detection unit, decision output unit, and post-processing module, respectively. It is used to control and coordinate the data transmission and working sequence between the modules to ensure that the entire computer image processing system operates in a predetermined manner. This includes, but is not limited to, instructing the preprocessing module to perform illumination correction and noise suppression of corresponding parameters based on the data acquired by the image acquisition module; guiding the semantic understanding unit to perform targeted image description generation based on the feature extraction results of the intelligent recognition module; and controlling the decision output unit to decide whether to generate a structured report based on the outlier identification results of the anomaly detection unit.

[0046] It should also include a display screen, connected to the controller, for displaying structured reports, etc., generated by the decision output unit.

[0047] In the feature extraction unit, MobileNetV3 employs depthwise separable convolution, inverse residual structure, hardware-friendly activation function h-swish, and SE module. It also optimizes the model structure through channel pruning, layer fusion, and adjustment of the width scaling factor. During training, QAT is used to simulate 8-bit integer calculations, and during deployment, the weights are quantized using INT8 and the activation values ​​are quantized using dynamic range quantization.

[0048] The semantic understanding unit adopts an encoder-decoder framework or a multimodal pre-trained model architecture. The visual feature extraction model includes one or more of ResNet-101, ViT-L / 16, CLIP-ViT, and Faster R-CNN. The text generation module adopts one of LSTM+Attention, Transformer, and GPT-2. Visual and text features are fused by calculating a cross-modal attention weight matrix.

[0049] The Autoencoder of the anomaly detection unit includes a basic Autoencoder structure and one or more improved variants of VAE, ConvAE, LSTM-AE, and Attention AE. During training, one or more of mean squared error, weighted MSE, adversarial loss, and Mahalanobis distance are used as loss functions. The anomaly determination threshold is selected by one or more of the following methods: quantile method, K-Sigma, extreme value theory, and dynamic threshold.

[0050] The post-processing module corrects misidentifications based on Conditional Random Field (CRF). It balances univariate and binary potentials by modeling the contextual dependencies between pixels / regions. The univariate potential is based on the initial classification confidence, while the binary potential is based on the similarity constraints of spatial features, color features, and semantic features.

[0051] Includes a support part 9, on both sides of which the two ends of a bracket 1 are fixed. The bracket 1 is arched. The upper and lower parts of the bracket 1 are provided with a fixed end of a lifting part 2. The movable end of the lifting part 2 is fixedly connected to a flattening component. The flattening component is used for flattening paper.

[0052] The support portion 9 is also connected to a scanning component, which is used to monitor any unflattened areas of the paper placed inside the support portion 9.

[0053] In this invention, the support part 9 is used to place the paper, and the vision sensor in the scanning component scans the paper to detect whether it is flattened. If an unflattened area is detected, the controller controls the lifting part 2 to move the flattening component downward. After the roller 701 comes into contact with the paper, the pressure sensor transmits a signal to the controller, and the lifting part 2 stops moving downward. The drive part 14 drives the drive rod 17 to swing. The drive rod 17 moves along the limiting slot 4, causing the connecting rod 5 to swing, which in turn causes the rotating plate 6 to rotate. The rotating plate 6 drives the remaining connecting rods 5 to swing, and the drive rod 17 drives the flattening part 7 to move outward and unfold. The roller 701 moves and rotates accordingly, realizing the flattening operation of the paper.

[0054] The flattening assembly includes a limiting part 3, a limiting slot 4, a connecting rod 5, a fixing plate 16, a rotating plate 6, a flattening part 7, a fixing shaft 8, and a driving part 14. The flattening part 7 consists of several sector-shaped blocks, all of which can be enclosed to form a circular plate with a central hole. The fixing plate 16 fixes the movable end of the lifting part 2. A positioning shaft 18 is fixedly connected to the lower center of the fixing plate 16. The positioning shaft 18 matches the central hole of the circular plate formed by all the sector-shaped blocks. The outer arc surface of the positioning shaft 18 can abut against the inner arc surface of the sector-shaped blocks. A rotating plate 6 is provided on the lower side of the fixing plate 16. The rotating plate 6 is circular and rotatably connected to the positioning shaft 18. The flattening part 7 is provided on the lower side of the rotating plate 6. One end of a drive rod 17 is fixedly connected to the middle part of the outer arc surface of the flattening part 7. The drive rod 17 is bent. The other end of the drive rod 17 is slidably disposed in the limiting groove 4. The limiting groove 4 is fixed at the upper edge of the fixed plate 16. Each drive rod 17 is rotatably connected to one end of the connecting rod 5. The other end of each connecting rod 5 is rotatably connected to the fixed shaft 8. The fixed shaft 8 is circumferentially fixed at the upper edge of the rotating plate 6. Rollers 701 are respectively provided at the lower end of the positioning shaft 18 and the lower side of each flattening part 7. The lower sides of the rollers 701 are flush. The drive part 14 is rotatably connected to the drive rod 17. The drive part 14 is rotatably disposed on the fixed plate 16.

[0055] In this embodiment, a pressure sensor is embedded in the roller 701 to monitor the contact between the roller 701 and the paper. When the roller 701 contacts the paper and the support part 9, the lifting part 2 stops moving downward, and the pressure sensor and the controller transmit data.

[0056] In this embodiment, the drive unit 14 can drive the drive rod 17 to swing, the drive rod 17 moves along the limiting slot 4, the drive rod 17 drives the connecting rod 5 to swing, the connecting rod 5 drives the rotating plate 6 to rotate, the rotating plate 6 drives the remaining connecting rods 5 to swing, the remaining connecting rods 5 drive the drive rod 17 to move along the limiting slot 4, the drive rod 17 drives the flattening part 7 to move outward and unfold, the roller 701 follows the movement and rotates during the movement, the rotation of the roller 701 reduces the friction with the paper, making it easier to move, the roller 701 follows the flattening part 7 to realize the flattening operation of the paper.

[0057] The drive unit 14 is a first electric push rod, and a protrusion is fixed at the other end of one of the drive rods 17. The protrusion is rotatably connected to the telescopic rod of the first electric push rod. The outer shell end of the first electric push rod is rotatably connected to the upper side of the fixed plate 16. Several circumferentially evenly arranged fixed shafts 8 are fixedly provided at the upper edge of the fixed plate 16. The fixed shafts 8 are set at the middle position when the flattening part 7 unfolds and moves outward.

[0058] In this embodiment, the telescopic rod of the first electric push rod can drive the drive rod 17 to move, and the first electric push rod swings during the movement of the drive rod 17.

[0059] There is at least one lifting part 2, which is an electric push rod. The outer shell of the second electric push rod is fixed to the bracket 1, and the telescopic rod of the second electric push rod is fixed to the center of the rotating plate 6.

[0060] In this embodiment, the extension and retraction of the telescopic rod of the second electric push rod can drive the flattening assembly to rise and fall.

[0061] Fixed blocks 10 are fixedly connected to both ends of one side of the support part 9. The fixed blocks 10 are rotatably connected to both ends of the connecting screw 11. The screw 11 is screwed to the moving block 12. The moving block 12 is fixedly connected to the center of the scanning part 15. The scanning part 15 is located on the upper side inside the support part 9. Several evenly arranged visual sensors are fixedly installed on the lower side of the scanning part. A motor 13 is fixed to one end of the screw 11. The output shaft of the motor 13 is fixedly connected to the fixed block 10. The visual sensor is the camera. The controller transmits data with the visual sensor. The controller is electrically connected to the lifting part 2, the driving part 14 and the motor 13.

[0062] In this embodiment, the paper is scanned by a scanning component. The motor 13 is a stepper motor. The rotation of the output shaft of the motor 13 can drive the screw 11 to rotate. The screw 11 can drive the moving block 12 to move. The moving block 12 drives the scanning unit 15 and the vision sensor to move, thereby realizing the scanning of the paper. The motor 13 drives the screw 11 to rotate forward for a period of time and then rotates in reverse. When the scanning unit 15 returns to its original position, the motor 13 stops rotating, thus realizing one reciprocating scanning process of the paper.

[0063] The method of using this invention is as follows:

[0064] Turn on the system power, initialize the controller and ensure that each module is working properly;

[0065] Place the image or paper to be processed in the appropriate position, and the image acquisition module will automatically acquire image data, while the scanning component will scan the paper.

[0066] The system automatically processes images in the order of preprocessing, intelligent recognition, decision output, and post-processing. During paper processing, the scanning component automatically scans the paper, detects any unflattened areas, and automatically controls the flattening component to flatten them if any are found.

[0067] After processing, users can view the structured report generated by the decision output unit.

[0068] The above-disclosed embodiments are merely specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A computer image processing system employing intelligent recognition technology, characterized in that, include: The image acquisition module includes a camera, a high-resolution sensor, an infrared / thermal imaging device, and a 3D LiDAR, and achieves low-latency data stream transmission through a 5G / fiber optic network. A preprocessing module is used for dynamic illumination correction and noise suppression based on adaptive filters; The intelligent recognition module includes a feature extraction unit, which uses a lightweight deep learning model combined with an attention mechanism to achieve feature extraction. The lightweight deep learning model is MobileNetV3, and the attention mechanism is ECA-Net or Dynamic Spatial Attention DSAM. The semantic understanding unit is used to generate image descriptions; Anomaly detection unit identifies outliers based on Autoencoder; Decision output unit, used to generate structured reports; The post-processing module corrects misidentifications based on contextual information. The controller is connected to the image acquisition module, preprocessing module, intelligent recognition module, semantic understanding unit, anomaly detection unit, decision output unit, and postprocessing module, respectively. It is used to control and coordinate the data transmission and working sequence between the modules to ensure that the entire computer image processing system operates in a predetermined manner. This includes, but is not limited to, instructing the preprocessing module to perform illumination correction and noise suppression of corresponding parameters based on the data acquired by the image acquisition module; guiding the semantic understanding unit to perform targeted image description generation based on the feature extraction results of the intelligent recognition module; and controlling the decision output unit to decide whether to generate a structured report based on the outlier identification results of the anomaly detection unit. It also includes a support part (9), on both sides of which the two ends of the bracket (1) are fixed. The bracket (1) is arched. The upper side of the bracket (1) is provided with a fixed end of a lifting part (2). The movable end of the lifting part (2) is fixedly connected to a flattening component. The flattening component is used for flattening paper. The support (9) is also connected to a scanning component, which is used to monitor the unflattened areas of the paper placed inside the support (9); The support part (9) has two fixed blocks (10) fixedly connected to one side. The fixed blocks (10) are rotatably connected to the two ends of the connecting screw (11). The screw (11) is screwed to the moving block (12). The moving block (12) is fixedly connected to the center of the scanning part (15). The scanning part (15) is located on the upper side inside the support part (9). Several evenly arranged visual sensors are fixedly arranged on the lower side of the scanning part. One end of the screw (11) is fixed to the motor (13). The output shaft of the motor (13) is fixedly connected to the fixed block (10). The visual sensor is the camera. The controller transmits data to the visual sensor. The controller is electrically connected to the lifting part (2), the driving part (14), and the motor (13). The flattening assembly includes a limiting part (3), a limiting slot (4), a connecting rod (5), a fixed plate (16), a rotating plate (6), a flattening part (7), a fixed shaft (8), and a driving part (14). The flattening part (7) consists of several sector blocks, all of which can be enclosed to form a circular plate with a central hole. The fixed plate (16) fixes the movable end of the lifting part (2). The lower center of the fixed plate (16) is fixedly connected to a positioning shaft (18), which matches the central hole of the circular plate formed by all the sector blocks. The outer arc surface of the positioning shaft (18) can abut against the inner arc surface of the sector blocks. A rotating plate (6) is provided on the lower side of the fixed plate (16). The rotating plate (6) is circular. The rotating plate (6) and the positioning shaft (18) are rotatably connected. Next, the lower side of the rotating plate (6) is provided with the flattening part (7). The middle part of the outer arc surface of the flattening part (7) is fixedly connected to one end of the driving rod (17). The driving rod (17) is bent. The other end of the driving rod (17) is slidably disposed in the limiting slot (4). The limiting slot (4) is fixed at the upper edge of the fixed plate (16). Each driving rod (17) is rotatably connected to one end of the connecting rod (5). The other end of each connecting rod (5) is rotatably connected to the fixed shaft (8). The fixed shaft (8) is circumferentially fixed at the upper edge of the rotating plate (6). The lower end of the positioning shaft (18) and the lower side of each flattening part (7) are respectively provided with rollers (701). The lower sides of the rollers (701) are flush.

2. The computer image processing system employing intelligent recognition technology according to claim 1, characterized in that, In the feature extraction unit, MobileNetV3 employs depthwise separable convolution, inverse residual structure, hardware-friendly activation function h-swish, and SE module. It also optimizes the model structure through channel pruning, layer fusion, and adjustment of the width scaling factor. During training, QAT is used to simulate 8-bit integer calculations, and during deployment, the weights are quantized using INT8 and the activation values ​​are quantized using dynamic range quantization.

3. The computer image processing system employing intelligent recognition technology according to claim 1, characterized in that, The semantic understanding unit adopts an encoder-decoder framework or a multimodal pre-trained model architecture. The visual feature extraction model includes one or more of ResNet-101, ViT-L / 16, CLIP-ViT, and Faster R-CNN. The text generation module adopts one of LSTM + Attention, Transformer, and GPT-2. Visual and text features are fused by calculating a cross-modal attention weight matrix.

4. The computer image processing system employing intelligent recognition technology according to claim 1, characterized in that, The Autoencoder of the anomaly detection unit includes a basic Autoencoder structure and one or more improved variants of VAE, ConvAE, LSTM-AE, and Attention AE. During training, one or more of mean squared error, weighted MSE, adversarial loss, and Mahalanobis distance are used as loss functions. The anomaly determination threshold is selected by one or more methods of quantile method, K-Sigma, extreme value theory, and dynamic threshold.

5. The computer image processing system employing intelligent recognition technology according to claim 1, characterized in that, The post-processing module corrects misidentifications based on Conditional Random Field (CRF). It balances univariate and binary potentials by modeling the contextual dependencies between pixels / regions. The univariate potential is based on the initial classification confidence, while the binary potential is based on the similarity constraints of spatial features, color features, and semantic features.

6. The computer image processing system employing intelligent recognition technology according to claim 1, characterized in that: One of the drive rods (17) has a fixed protrusion at the other end. The protrusion is rotatably connected to the telescopic rod of the first electric push rod. The outer shell end of the first electric push rod is rotatably connected to the upper side of the fixed disk (16). Several fixed shafts (8) are evenly arranged in a circle at the upper edge of the fixed disk (16). The fixed shafts (8) are set at the middle position when the flattened part (7) unfolds and moves outward.

7. The computer image processing system employing intelligent recognition technology according to claim 6, characterized in that: There is at least one lifting part (2), which is an electric push rod. The outer shell of the second electric push rod is fixed to the bracket (1), and the telescopic rod of the second electric push rod is fixed to the center of the rotating plate (6).

Images