Detection method and system for warehouse-in and warehouse-out of non-inductive warehouse goods

Abstract

The invention discloses a detection method and system for non-inductive warehouse goods in and out of a warehouse, and the method is characterized in that the method comprises the following steps: collecting image information of an electric power fitting warehouse, and obtaining a to-be-detected image through the work of image labeling, image preprocessing, image enhancement and the like; analyzing the to-be-measured image through a polygon approximation method, and calculating the contour of the to-be-measured target; extracting a feature map by using a neural network, and identifying a to-be-detected target by using a coding and decoding method so as to obtain a plurality of parallax transformation modules; through the output of the plurality of parallax conversion modules, the to-be-detected image carries out non-inductive detection on the goods in the electric power fitting warehouse.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and more specifically, to a method and system for detecting the entry and exit of goods in a non-contact warehouse. Background Art

[0002] In modern warehouse management, traditional inbound and outbound management systems usually rely on manual operations, including barcode scanning, manual data entry, and regular inventory. Although this method can meet basic needs to a certain extent, its limitations gradually become apparent as the scale of warehousing expands. Specifically, traditional methods have many disadvantages, such as frequent human errors, low work efficiency, data lag, and high management costs. In addition, with the rapid development of e-commerce and logistics industries, the requirements for the speed of goods circulation and the accuracy of warehouse management are constantly increasing, and the traditional inbound and outbound management model can no longer meet the needs of modern warehousing systems.

[0003] The development of computer vision technology has provided a new opportunity to solve the above problems. Computer vision is a technology that automatically processes and analyzes images or videos through computers, which can extract useful information from them and perform further processing and analysis. In the field of warehouse management, the application of computer vision technology can realize the automatic identification, positioning, tracking and status monitoring of goods, thereby greatly improving the efficiency and accuracy of inbound and outbound management. By deploying smart cameras and sensors, combined with deep learning and image processing algorithms, computer vision systems can realize real-time monitoring, automatic identification and status recording of goods, greatly reducing dependence on manpower.

[0004] The contactless warehouse in-and-out management system is a new type of warehouse management system based on computer vision technology. The system deploys cameras and sensors in the warehouse and uses computer vision algorithms to monitor and automatically manage the in-and-out process of goods in real time, thus realizing contactless intelligent management. Specifically, the contactless warehouse in-and-out management system can automatically identify the type, quantity and location of goods, automatically generate in-and-out records, update inventory data in real time, and promptly alarm when abnormal situations occur.

[0005] In this system, cameras and sensors are distributed in every corner of the warehouse to form a fully covered monitoring network. When goods enter or leave the warehouse, the system automatically captures images or videos, and analyzes and processes them through computer vision algorithms to identify the type, quantity and location of the goods. At the same time, the system automatically generates in and out of the warehouse records, updates inventory data, and uploads the data to the central database. The database can be seamlessly connected with other management systems of the enterprise (such as ERP system) to achieve real-time data sharing and collaborative management.

[0006] This non-sensitive management mode can not only significantly reduce manual operations and improve work efficiency, but also avoid errors and delays caused by human factors. At the same time, through real-time monitoring and data analysis, the system can provide more accurate inventory management and decision support, further optimize the utilization of storage resources, and reduce operating costs. For example, through the analysis of cargo flow data, enterprises can optimize the storage layout, reduce the distance of cargo transportation, and improve the utilization of storage space. In addition, the system can also predict future inventory needs through the analysis of historical data, helping enterprises to formulate more scientific procurement and production plans.

[0007] With the continuous advancement of artificial intelligence and Internet of Things technologies, the non-contact warehouse in-and-out management system will play an increasingly important role in future warehouse management, driving the entire industry towards intelligence and automation. In the future, the system can also be combined with robotics and drone technology to further enhance the intelligence level of warehouse management and achieve fully automated warehouse management. For example, robots can automatically carry and place goods under the guidance of the system, and drones can conduct inspections and inventory inside the warehouse, thereby further reducing manpower input and improving management efficiency.

[0008] In summary, the non-contact warehouse in-and-out management system has achieved a revolutionary change in the traditional warehouse management model by introducing computer vision technology. The system not only improves work efficiency and data accuracy, but also significantly reduces management costs, bringing significant economic benefits and competitive advantages to enterprises. With the continuous development and improvement of technology, the non-contact warehouse in-and-out management system will play a more important role in future warehouse management and become one of the core technologies of modern warehouse management.

[0009] Computer vision technology is a technology that automatically processes and analyzes images or videos through computers to extract useful information from them. Specifically for the non-contact warehouse in-and-out management system, the application of computer vision technology includes the following aspects: Image capture and processing: By arranging high-resolution cameras in the warehouse, images or videos of goods entering and leaving the warehouse are captured in real time. Using image processing algorithms, the captured images are preprocessed, such as denoising, enhancement, correction, etc., to improve image quality and recognition accuracy. Object recognition and classification: Deep learning algorithms, especially convolutional neural networks (convolution), are used to identify and classify goods in images. The algorithm can automatically learn and extract the characteristics of goods, such as shape, color, labels, etc., so as to realize automatic recognition of the type of goods. Object tracking and positioning: Using multi-target tracking algorithms and 3D reconstruction technology, the movement trajectory of goods in the warehouse is tracked and their position in the warehouse is determined in real time. These algorithms can handle complex scenes, such as occlusion, overlap, etc., to ensure the accurate positioning of goods. State monitoring and anomaly detection: Through computer vision technology, the state of goods is monitored in real time, such as whether the packaging is intact or damaged. Combined with machine learning algorithms, it can detect and warn of abnormal situations, such as lost goods and abnormal handling.

[0010] Sensor technology also plays a key role in the non-contact warehousing in-and-out management system. The main sensors include: RFID sensors: By installing RFID tags on the goods, automatic identification and tracking of goods can be achieved. RFID sensors can read the label information of the goods in real time and upload the data to the central database. LiDAR: Using LiDAR technology, the warehouse environment is scanned in three dimensions to build a three-dimensional model of the warehouse. LiDAR can accurately measure the distance and position of the goods, and assist computer vision in cargo positioning and path planning. Environmental sensors: including temperature sensors, humidity sensors, smoke sensors, etc., are used to monitor changes in the warehouse environment, ensure that the environment in the warehouse is suitable for cargo storage, and prevent safety hazards such as fire.

[0011] The contactless warehouse in-and-out management system needs to process and analyze a large amount of data. The technologies applied mainly include: Big data processing technology: Through distributed computing frameworks (such as Hadoop, Spark, etc.), the sensor data, image data, RFID data, etc. in the warehouse are efficiently processed and stored to achieve real-time processing and analysis of massive data. Artificial intelligence and machine learning technology: Using machine learning algorithms, historical data is analyzed and modeled to achieve forecasts of inventory demand, optimization of warehouse layout, detection of abnormal situations, etc. Through continuous learning and optimization, the system can improve the accuracy and efficiency of decision-making. Internet of Things technology: Through the Internet of Things platform, various sensors, cameras, robots and other equipment can be interconnected to achieve real-time data collection, transmission and processing. Internet of Things technology can improve the intelligence level of the system and enhance the system's adaptive ability.

[0012] The future warehouse management system will be more intelligent. By introducing artificial intelligence technology, the system can achieve self-learning and self-adaptation, and can adjust its own working mode according to environmental changes and needs. For example, the intelligent warehouse system can optimize the warehouse layout, adjust the cargo storage strategy, and predict inventory demand based on historical data and real-time monitoring data, thereby improving the efficiency and accuracy of warehouse management. Automation is one of the development directions of warehouse management in the future. By introducing robot technology, drone technology, etc., all aspects of warehouse management (such as cargo handling, placement, inventory, etc.) will be fully automated. Robots can automatically complete the handling and placement of goods under the guidance of the system, and drones can conduct inspections and inventory inside the warehouse, thereby reducing manpower input and improving management efficiency. The development of Internet of Things technology enables various devices of the warehouse management system (such as sensors, cameras, robots, etc.) to be interconnected and form an overall collaborative intelligent system. Through the Internet of Things platform, various devices can share data in real time, work together, and improve the response speed and coordination ability of the system. For example, when the sensor detects an abnormal situation, the system can automatically dispatch cameras for monitoring and robots for processing to ensure the safety and efficiency of warehouse management. Warehouse management in the future will be more data-driven. Through the processing and analysis of massive data, the system can provide more accurate decision support. Big data analysis technology will be widely used in warehouse management to achieve the prediction of inventory demand, optimization of warehouse layout, detection of abnormal situations, etc. Data-driven warehouse management will improve the accuracy and efficiency of decision-making, reduce operating costs, and enhance the competitiveness of enterprises. With the intelligence and automation of warehouse management systems, the security of the system will become particularly important. The warehouse management system in the future will introduce more advanced security technologies, such as blockchain technology, encryption technology, etc., to ensure data security and privacy protection. At the same time, through multi-level security protection measures, the system can effectively respond to various security threats and ensure the security and reliability of warehouse management.

[0013] In summary, the computer vision-based non-contact warehouse in-and-out management system is gradually realizing intelligent, automated, interconnected, data-driven and highly secure warehouse management by introducing advanced computer vision technology, sensor technology, data processing and analysis technology. With the continuous development and improvement of technology, this new warehouse management system will bring significant economic benefits and competitive advantages to enterprises, and promote the entire industry to develop in the direction of intelligence and automation.

[0014] At present, in the field of intelligent warehousing, material management is mostly based on RFID technology. In the process of warehousing management, warehousing procurement, inbound and outbound quality verification, ledger entry and other tasks are all completed by manual identification, pasting RFID tags and entry. Warehouse materials are of various types and models. First of all, it is difficult for existing manual identification and RFID technology to detect small differences and quality problems of materials, and it is easy to make mistakes in picking materials and inconsistent quality of inbound and outbound materials. These defects will bring serious safety hazards; secondly, the different business proficiency of operators leads to errors in pasting RFID tags, and there are differences between actual inventory and ledger records. The work efficiency is not high, which is contrary to the policy of central enterprises to reduce costs and increase efficiency; thirdly, due to the reverse reflective characteristics of ultra-high frequency RFID electronic tags, it is difficult to use them in commodities such as metals, and the price of RFID electronic tags is higher than that of ordinary barcode tags, which is dozens of times that of ordinary barcode tags. If the usage is large, the cost will be very high. Based on the above situation, how to ensure the quality of materials entering and leaving the warehouse, as well as the accuracy, authenticity, completeness and efficiency of ledger records, reduce the cost of post-verification, and improve grassroots work efficiency are urgent issues facing smart warehousing.

[0015] Computer vision technology based on artificial intelligence is combined with images of warehouse materials and equipment, historical ledger information and data identification information to build an intelligent warehousing image recognition system. It is combined with existing RFID technology to realize intelligent material identification, intelligent equipment management, and visual knowledge management. It covers business links such as material entry and exit, regular sampling, and acceptance. It solves problems such as differences between inventory materials and system ledgers, and low employee operation efficiency. It improves employee business skills, improves the accuracy of material entry and exit, improves overall work efficiency, and reduces enterprise operating costs.

[0016] However, existing methods have problems, such as: Low efficiency: Traditional warehouse management systems rely heavily on manual operations, including barcode scanning, manual data entry, and regular inventory. These operation steps are cumbersome and time-consuming, especially when handling large quantities of goods, and the efficiency is extremely low. For example, in large warehouses, relying solely on manual scanning and data entry requires a lot of time and manpower, resulting in a significant decrease in overall work efficiency. Frequent errors: Manual operation processes are prone to errors, such as errors in barcode scanning of goods, data entry errors, and errors in classification of goods. These human errors not only affect the accuracy of inventory records, but may also lead to problems such as loss, misdelivery, and delays in goods, further affecting the reliability and efficiency of warehouse management. Insufficient mechanization: Although some warehouse systems have introduced some automated equipment, such as automated stereoscopic warehouses and conveyor belt systems, the overall degree of automation is still low. Many key links, such as the handling, entry and exit of goods, still require a lot of manual intervention, and the automation of the entire process cannot be achieved. Low system integration: Existing automated equipment usually operates independently, lacks unified management and coordination, resulting in low overall system operation efficiency. For example, the data between the automated warehouse system and the inventory management system cannot be shared in real time, resulting in information islands and affecting overall management efficiency. Poor real-time performance and data lag: The data of the traditional warehouse management system is not updated in a timely manner, and the inventory information cannot reflect the actual situation in real time. Due to the lag of manual inventory and data entry, there is often a certain delay in inventory information, which cannot accurately reflect the current inventory status, and easily leads to excess or shortage of inventory. Slow response speed: Due to the slow data processing and transmission speed, it is difficult for traditional systems to monitor the status and location of goods in real time, and it is impossible to detect and handle abnormal situations in time. For example, when the goods are lost or damaged, the system cannot alarm in time, resulting in the problem not being solved in time, affecting the effectiveness of warehouse management. Poor flexibility and difficulty in system expansion: The existing warehouse management system is usually fixed, and it is difficult to flexibly expand and adjust according to demand, and cannot adapt to the rapidly changing market demand. When enterprises expand the scale of storage or adjust management strategies, they need to reconfigure or replace the system, which is costly and time-consuming. Lack of intelligent analysis: Traditional systems lack intelligent data analysis and decision support functions, and cannot optimize storage layout and management strategies through data analysis. For example, the system cannot automatically adjust the storage location and order of goods in and out of the warehouse based on historical data and real-time monitoring data, which reduces the flexibility and efficiency of warehouse management.

[0017] In view of the above problems, there is an urgent need for a detection method and system for the entry and exit of goods in a contactless warehouse. Summary of the invention

[0018] In order to solve the deficiencies in the prior art, the present invention provides a method and system for detecting the entry and exit of goods in a contactless warehouse.

[0019] The present invention adopts the following technical solution.

[0020] The first aspect of the present invention relates to a method for detecting the entry and exit of goods in a contactless warehouse, and the method comprises the following steps: collecting image information of a power parts warehouse, and obtaining an image to be tested through image labeling, image preprocessing, image enhancement, etc.; analyzing the image to be tested through a polygonal approximation method, and calculating the contour of the target to be tested; applying a neural network to extract a feature map, using a coding and decoding method to identify the target to be tested, and obtaining multiple parallax transformation modules based on this; and implementing contactless detection of goods in the power parts warehouse through the output of multiple parallax transformation modules and the image to be tested.

[0021] Preferably, image information of a power parts warehouse is collected, and images to be tested are obtained through image labeling, image preprocessing, image enhancement, etc., including: manually labeling the images of power parts to obtain the product names of the images of power parts; image preprocessing includes first determining the effective area of ​​the electronic image after grayscale transformation and binarization, and then drawing the minimum circumscribed rectangle of the effective area, and correcting the original electronic image by calculating the deflection angle of the rectangle to solve the problem of tilt and distortion; image enhancement obtains the rotation matrix of the image through the getRotationMatrix2D function, and obtains the rotated image through the affine transformation function.

[0022] Preferably, the image to be measured is analyzed by a polygonal approximation method to calculate the contour of the target to be measured, including: filtering the image by a Gaussian filtering method; dividing the image grayscale histogram into two parts by using an optimal threshold value so that the variance between the two parts is maximized to separate the target from the background.

[0023] Preferably, a chord of the target curve is made, assuming that the first point of the curve is a and the last point is b, the first point is connected to the last point to form a straight line AB as the chord of the target curve; the maximum distance point C between the target curve and the chord is calculated, so that the maximum distance between the curve and the chord is d; d is compared with a preset threshold, if it is less than the specified threshold, the straight line segment can be approximated as the target curve, and the processing of the curve segment is completed; if the distance d is greater than the threshold, the line segment is divided into two line segments AC and BC, and C is used as the point, and the above operation is continued until all curves are processed; a polygon is fitted by the Douglas-Peucker algorithm, and a polygon approximation algorithm is used to identify the shape of the object.

[0024] Preferably, a neural network is applied to extract feature maps, and a coding and decoding method is used to identify the target to be measured, thereby obtaining multiple parallax transformation modules, including: the neural network involves convolution, activation function, pooling operation and out module; taking a two-dimensional image signal I as input, sliding a frame of the same size and convolution kernel from the upper left corner to the lower right corner of the input image; checking the multiplication and addition results of the corresponding elements each time the selected image area is slid and convolution is performed; calculating a new value S(i, j) in the input activation function f(S(i, j)), and the activation function selects a Sigmoid function, a hyperbolic tangent function or a rectified linear unit function; reducing the input size of the next layer, reducing the amount of calculation and the number of parameters, and obtaining maximum pooling; when training the neural network, randomly selecting and ignoring some hidden layer nodes with a certain probability.

[0025] Preferably, a neural network is applied to extract feature maps, and a coding and decoding method is used to identify the target to be measured, thereby obtaining multiple disparity transformation modules, including: removing the convolution 10_2 in the original SSD network and the features obtained from the convolution 9_2; using the basic network VGG16, adding convolution 6, convolution 7, convolution 8_2 and convolution 9_2 in sequence; obtaining four different scale features of VGG16, convolution 7, convolution 8_2, and convolution 9_2 respectively to predict the position of the target bounding box and the corresponding confidence level.

[0026] Preferably, the output of multiple disparity transformation modules and the image to be tested are used to implement non-sensing detection of goods in the power parts warehouse, including: obtaining a target recognition network model based on an encoder-decoder architecture; the encoder is responsible for extracting semantic information and contextual features from the input image, and the decoder maps these features back to the original resolution to generate the final segmentation result.

[0027] The second aspect of the present invention relates to a detection system for the entry and exit of goods in a contactless warehouse using the method of the first aspect of the present invention; the system includes an acquisition module, an analysis module, a transformation module, and a detection module; wherein the acquisition module is used to collect image information of a power parts warehouse, and obtain a test image through image labeling, image preprocessing, image enhancement, etc.; the analysis module is used to analyze the test image through a polygonal approximation method to calculate the contour of the target to be tested; the transformation module is used to apply a neural network to extract a feature map, and use a coding and decoding method to identify the target to be tested, thereby obtaining multiple parallax transformation modules; the detection module is used to implement contactless detection of goods in the power parts warehouse through the output of multiple parallax transformation modules and the test image.

[0028] A third aspect of the present invention relates to a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0029] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect of the present invention.

[0030] The beneficial effect of the present invention lies in that, compared with the prior art, the present invention is a detection method for the entry and exit of goods in a non-contact warehouse, characterized in that the method comprises the following steps: collecting image information of the power parts warehouse, obtaining the image to be tested through image labeling, image preprocessing, image enhancement and the like; analyzing the image to be tested by a polygonal approximation method, and calculating the contour of the target to be tested; applying a neural network to extract feature maps, using a coding and decoding method to identify the target to be tested, and obtaining multiple parallax transformation modules based on this; and implementing non-contact detection of goods in the power parts warehouse through the output of multiple parallax transformation modules and the image to be tested.

[0031] The beneficial effects of the present invention also include:

[0032] 1. Efficient automation, automatic identification and recording: Using computer vision and RFID technology, the system can automatically identify the type, quantity and location of goods, automatically generate inbound and outbound records, and reduce manual operations. For example, when goods enter or leave the warehouse, smart cameras and RFID readers will automatically capture and record the information of the goods, and upload the data to the central database in real time to achieve automated inbound and outbound records.

[0033] 2. Real-time monitoring and tracking: Through smart cameras and sensors, the system can monitor the status and location of goods in real time and accurately track the movement of goods. For example, the system can use multi-target tracking algorithms to track every item in the warehouse in real time to ensure that its location and status are accurately recorded to avoid loss or misplacement of goods.

[0034] 3. Intelligent management, abnormality detection and alarm: The system uses computer vision and machine learning algorithms to monitor the status of goods and environmental parameters in real time, detect abnormal situations and issue alarms in a timely manner. For example, when the system detects that a certain item has been moved abnormally or the packaging is damaged, it will immediately issue an alarm and send the abnormal information to the relevant management personnel to ensure that the problem is handled in a timely manner.

[0035] 4. Data analysis and decision support: The system can analyze and model historical data, predict inventory demand, and optimize warehouse layout. For example, the system can analyze historical sales data and seasonal changes to predict future inventory demand, arrange procurement plans reasonably, and avoid excess or shortage of inventory. At the same time, the system can optimize the storage location of goods based on the flow data of goods, reduce the transportation distance, and improve the utilization rate of storage space.

[0036] 5. High integration and coordination, system integration: The non-contact warehousing and warehousing management system can be seamlessly connected with the enterprise's ERP system and other management systems to achieve real-time data sharing and collaborative management. For example, the system can automatically synchronize real-time updated inventory data to the ERP system to ensure that all departments can obtain the latest inventory information and improve the overall operational efficiency of the enterprise.

[0037] 6. Equipment interconnection: Through the Internet of Things technology, various devices (such as cameras, sensors, robots, etc.) are interconnected to form an overall coordinated intelligent system. For example, when the environmental sensor detects that the temperature or humidity in the warehouse exceeds the set range, the system can automatically adjust the air conditioning or humidification equipment to ensure that the cargo storage environment is suitable.

[0038] 7. Flexible scalability and modular design: The system adopts modular design and can be flexibly expanded and adjusted according to needs. For example, enterprises can add or reduce cameras, sensors, RFID readers and other equipment according to the storage scale and management needs, easily expand system functions, and adapt to different storage environments and business needs.

[0039] 8. Adaptive optimization: Through continuous learning and optimization, the system can adjust its working mode according to environmental changes and needs. For example, the system can automatically adjust the camera's shooting angle and focal length by analyzing real-time monitoring data, improve monitoring coverage and clarity, and enhance the system's adaptive ability.

[0040] 9. Safety, reliability and environmental monitoring: The system uses environmental sensors to monitor the temperature, humidity and other environmental parameters in the warehouse in real time to prevent potential safety hazards such as fire. For example, when the system detects that the smoke concentration in the warehouse exceeds the standard, it will immediately issue a fire alarm and activate the automatic fire extinguishing system to ensure the safety of the warehouse.

[0041] 10. Multi-level security protection: The system adopts multi-level security protection measures, such as data encryption and identity authentication, to ensure data security and privacy protection. For example, the system uses encryption technology during data transmission to prevent data from being stolen or tampered with; at the same time, the system uses an identity authentication mechanism to ensure that only authorized personnel can access and operate the system to prevent illegal operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of convolution in the prior art;

[0043] Figure 2 It is a schematic diagram of maximum pooling in the prior art;

[0044] Figure 3 It is a schematic diagram of dropout in the prior art;

[0045] Figure 4 This is a detection model diagram of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer and more accurate, the technical solution of the present invention is described in detail below through multiple specific implementation methods. The embodiments adopted by the present invention are only used to explain the present invention and are not used to limit the content of the present invention.

[0047] The first aspect of the present invention relates to a method for detecting the entry and exit of goods in a contactless warehouse, characterized in that the method comprises the following steps: collecting image information of a power parts warehouse, and obtaining a waiting image through image labeling, image preprocessing, image enhancement and the like; analyzing the waiting image through a polygonal approximation method, and calculating the contour of the waiting target; applying a neural network to extract a feature map, using a coding and decoding method to identify the waiting target, and obtaining a plurality of parallax transformation modules based on this; and implementing contactless detection of goods in the power parts warehouse through the output of the plurality of parallax transformation modules and the waiting image.

[0048] The system deploys high-definition cameras and sensors inside and around the warehouse, covering key locations such as entrances, exits, shelf areas, and intersections. The cameras capture real-time video streams and transmit them to a central processing unit (usually the cloud or a local server) through an optimized network architecture. Sensors such as motion detectors and temperature sensors enhance the system's ability to perceive environmental changes and abnormal conditions.

[0049] The central processing unit uses computer vision technology to perform real-time image processing on the transmitted video stream. Deep learning algorithms such as YOLO (You Only Look Once) or SSD (Single Shot MultiBox Detector) are used for object detection and tracking. Detected objects are accurately identified, classified, and recorded in the database, including location, quantity, and other key attributes.

[0050] The system automatically performs inbound and outbound operations based on the identified and classified data, reducing human intervention. Automated equipment such as AGVs (automatic guided vehicles) and robots can be integrated with the system to achieve automatic sorting, handling and loading of goods. The system monitors abnormal situations and automatically triggers alarms, such as cargo retention or abnormal movement, to ensure timely processing and safety control.

[0051] Managers can understand the warehouse operation status through real-time feedback and detailed data reports from the system. The reports include detailed information such as inventory status, transportation efficiency, abnormal event records, etc., to support daily decision-making and long-term strategic planning.

[0052] Image information of the power parts warehouse is collected, and the images to be tested are obtained through image labeling, image preprocessing, image enhancement and other tasks, including: manually labeling the power parts images to obtain the product names of the power parts images; image preprocessing includes first determining the effective area of ​​the electronic image through grayscale transformation and binarization, and then drawing the minimum circumscribed rectangle of the effective area, and correcting the original electronic image by calculating the deflection angle of the rectangle to solve the problem of tilt and distortion; image enhancement obtains the image rotation matrix through the getRotationMatrix2D function, and obtains the rotated image through the affine transformation function.

[0053] After converting power accessories into electronic images through high-speed scanners and other equipment, a series of image processing operations are required. Data preparation is the first step in power accessory recognition, which determines the data quality of the input model, including image labeling, image preprocessing, image enhancement, etc. Image labeling is a unique step in model training on the training side. It is a process of manually annotating images of power accessories, such as labeling them with "miniature circuit breaker_DZ47-60_3P_C32", in order to help the machine learn and recognize the features in the data.

[0054] Image preprocessing mainly includes grayscale processing, binarization, and solving image tilt. Since the captured electronic images often have certain tilt and distortion, the electronic images need to be corrected first. After grayscale transformation and binarization, the effective area of ​​the electronic image is first determined, and then the minimum circumscribed rectangle of the effective area is drawn. By calculating the deflection angle of the rectangle, the original electronic image can be corrected to solve problems such as tilt and distortion. Image enhancement is also a unique step in model training at the training end, the purpose is to make limited data generate more data. Usually, the number and diversity of training samples are increased through operations such as rotation, translation, and cutout, noise data is increased, the model's dependence on certain attributes is reduced, and the robustness and generalization ability of the model are improved. The present invention obtains the rotation matrix of the image through the getRotationMatrix2D function, obtains the rotated image through the affine transformation function, and selects angle values ​​such as -30°, -15°, 15°, and 30°.

[0055] Image preprocessing is an indispensable step in object shape recognition and is the actual image acquisition. There are many external environmental interferences. In order to eliminate the influence of irrelevant information, the image needs to be preprocessed. The present invention uses a Gaussian filtering method to filter the image. In order to improve the reliability of target shape recognition, the binary segmentation method uses the maximum inter-class variance algorithm to separate the target from the background. The maximum inter-class variance algorithm is the best threshold selection algorithm in real image segmentation. The algorithm is simple to calculate and is not affected by image brightness and contrast. It can better separate the target from the background, thereby improving the accuracy of later target shape recognition. Other types of threshold segmentation algorithms are user-defined thresholds, while the maximum inter-class variance threshold type automatically obtains a threshold suitable for binary binarization through the maximum inter-class variance algorithm, so this method is more suitable for application in engineering practice. The maximum inter-class variance threshold type automatically obtains a threshold suitable for binary binarization through the maximum inter-class variance algorithm. The principle of the maximum inter-class variance algorithm is that the optimal threshold divides the image grayscale histogram into two parts, so that the variance between the two parts is maximized, that is, the separation is maximized. The relationship between the cumulative probability, average grayscale, and variance and the values ​​listed in the formula is as follows:

[0056]

[0057] The parameters in the above formula are commonly used in the maximum between-class variance.

[0058] Image feature extraction uses polygonal approximation to obtain the basic outline of the target. Polygonal approximation mainly identifies the shape of an object through a polygonal fitting function, and the polygonal fitting function is fitted by the Douglas-Peucker algorithm. The steps of applying the algorithm are as follows. First, make the chord of the target curve. Assume that the first point of the curve is a and the last point is b. Connect the first point with the last point to form a straight line AB as the chord of the target curve. Then, find the maximum distance point C between the target curve and the chord, so that the maximum distance between the curve and the chord is d. Compare d with a preset threshold. If it is less than the specified threshold, this section of the straight line can be approximated as the target curve, and the processing of this section of the curve is completed. If the distance d is greater than the threshold, the line segment is divided into two line segments AC and BC, and C is used as the point to continue the above operation until all curves are processed. The present invention defines the output polygon point set separately as the side length of the polygon, thereby determining the shape of the polygon. The grayscale centroid method is used to calculate the center of gravity of the object and locate the position of the moving target. The pixel points of the image are written as density function f(x, y), the expectation of the pixel point is the origin moment of the image, and its centroid coordinates are obtained by calculating the image moment.

[0059] Convolution is a special artificial neural network that applies the convolution algorithm to the neural network. It is the main method of the current DL algorithm and the main tool for studying artificial intelligence machine vision. It mainly involves four main parts of technology: convolution, activation function, pooling operation and technology. The form of the convolution operation described in the present invention is to use a two-dimensional (2D) image signal I as input, corresponding to the two-dimensional convolution kernel K, and the two-dimensional convolution operation process is as follows Figure 1 As shown. Then slide a box of the same size and convolution kernel from the upper left corner of the input image to the lower right corner. Each time you slide the selected image area and perform convolution, check the multiplication and addition results of the corresponding elements. Figure 1 The convolution shown is the result of computing the rightmost matrix.

[0060] In convolution, it is necessary to calculate the new value S(i, j) in the input activation function f(S(i, j)). The activation function can generally be selected from the Sigmoid function, the hyperbolic tangent function, and the rectified linear unit function.

[0061] Sigmoid function is a real number input, which can be mapped to (0, 1), and its derivatives are all <0.25, especially the larger the absolute value of the input value, the smaller the derivative value, the phenomenon of gradient vanishing occurs in the multilayer neural network during training, that is, the gradient value is close to zero, and the BP network parameters cannot be updated, and the network cannot reach the optimal parameters. The advantage of the tangent function of the hyperbolic tangent function relative to the Sigmoid function is that the output mean is 0, but when the input absolute value is large, the gradient vanishing problem will occur. The rectified linear unit function is widely used as an activation function in recent years, because it is simple to calculate and converges quickly. Therefore, the convolution activation function rectified linear unit function is selected in the present invention.

[0062] Generally speaking, the pooling operation in convolution mainly includes maximum pooling and average pooling operations. Maximum pooling is the maximum value of the area, and average pooling is the average value of the area. The role of the pooling operation is to reduce the input size of the next layer, reduce the amount of calculation and the number of parameters, improve robustness, expand the perception area, and prevent overfitting. Before this, it was more common to use the average pooling operation, but in recent years, some scholars have shown that the use of maximum pooling can make the network performance better. Therefore, in the present invention, the maximum pooling operation is used in the network to replace the average pooling operation. The principle of the maximum pooling operation is as follows Figure 2 shown.

[0063] Dropout technology refers to randomly selecting and ignoring some hidden layer nodes with a certain probability when training a neural network. Figure 3As shown, using the dropout layer, some neurons may not participate in the operation. In the present invention, the probability value of the hidden layer node Dropout is 0.3. The application of the Dropout technology can become a different network model for each training. Through a large amount of training, it can be regarded as an average method of the network model. From the principle of the above Dropout technology, it can be seen that its main function is to avoid overfitting of the network and improve the generalization ability of the network, especially the popular convolution, which has a very good effect.

[0064] The basic structure of pallet detection convolution is as follows Figure 4 As shown, it mainly refers to the single-shot multi-box detector (SSD) detection algorithm. The improvement of the network structure in the present invention is mainly to directly remove the convolution 10_2 in the original SSD network and the features obtained from the convolution 9_2, and considering that the pap layers in the convolution 10_2 and the original SSD model are much smaller than the pap layers obtained in the previous layers, so that there is a serious imbalance when they are removed.

[0065] In the network model, the first part is the base network VGG16, but it will be used to classify all the connected layers after removing VGG16. In addition, VGG16 is followed by convolution layers of convolution 6, convolution 7, convolution 8_2 (including two layers of convolution operations of 1×1×256 and 3×3×512-s2), and convolution 9_2 (including two layers of convolution operations of 1×1×128 and 3×3×256-s2).

[0066] In the prediction stage, four different scale features, VGG16, Conv7, Conv8_2, and Conv9_2, are obtained to predict the location of the target bounding box and the corresponding confidence level. Unlike Overfeat and YOLO, only a single-scale feature map is used. It can be foreseen that using feature maps of different scales instead of using input images of different scales can improve the scale invariance of the network.

[0067] By utilizing feature maps of multiple scales, this method not only improves the detection capability of objects of different sizes, but also significantly enhances the robustness and accuracy of the model. This is because feature maps at different levels can capture different levels of visual information: feature maps at lower levels tend to capture fine-grained local features, while feature maps at higher levels are better at extracting global structural information. This multi-level feature fusion method enables the model to consider both local details and global context in a single forward propagation process, thereby effectively improving the accuracy of small object detection and the recall rate of large object detection.

[0068] The present invention also proposes a target recognition network model based on an encoder-decoder architecture. The encoder is responsible for extracting semantic information and contextual features from the input image, and the decoder maps these features back to the original resolution to generate the final segmentation result. In order to improve the expressiveness of the features, we introduced a multi-scale feature alignment module in the decoder. This module achieves effective fusion of bottom-level and high-level features by aligning the feature maps fused at different levels of the encoder with the corresponding feature maps in the decoder. This design aims to preserve the details and local features of the image and improve the perception range and semantic understanding ability of the model in the segmentation task. The decoder also includes an upsampling layer, a batch normalization layer, and a ReLU activation function layer. The upsampling layer is used to restore the resolution of the feature map and reduce the number of channels, thereby avoiding excessive computational complexity. The batch normalization layer and the ReLU activation function further enhance the feature expression and nonlinear modeling capabilities. The last layer generates a probability map by adding a softmax layer, and the probability value of each pixel can be used to obtain the final target recognition result by setting an appropriate threshold.

[0069] Through the output of multiple disparity transformation modules and the image to be tested, the goods in the power parts warehouse are non-sensingly detected, including: obtaining a target recognition network model based on an encoder-decoder architecture; the encoder is responsible for extracting semantic information and context features from the input image, and the decoder maps these features back to the original resolution to generate the final segmentation result;

[0070] The present invention may include one or more disparity transformation modules. This module may perform various operations on the image to be tested, such as the various operation methods mentioned above, thereby obtaining multiple similar disparity maps;

[0071] The present invention adopts a federated learning algorithm to use multiple disparity maps as sub-models and the original image to be tested as the main model, and finally obtains the optimized target detection result by iteratively training the parameters of a main model and multiple sub-models;

[0072] Multiple disparity information is highly correlated with the local features of the specific scene, while the RGB image of the original image to be tested often contains more global context information, which helps to understand the scene as a whole. Therefore, they are used as sub-models and main models respectively. This main model-sub-model combination can effectively balance local details and global consistency, and use the attention mechanism as an adaptive module to generate the weight of each feature. The adaptive module dynamically generates weights based on the distribution characteristics and context information of the input features. The generated weights are used to perform weighted operations on the features extracted by the main model and sub-model.

[0073] In the low-loss subspace, the features of different encoders are fused together through a smooth weighting strategy, so that they complement each other in the shared space. During model training, the weight adjustment module is continuously optimized to reduce the loss of information during feature fusion. Through the feedback of the loss function and the optimization algorithm, the weight of each feature is gradually adjusted so that the model can minimize the loss in the joint feature space.

[0074] A specific implementation is as follows:

[0075] Construct a mixed weight a, which is randomly selected from a distribution function U(0, 1). Construct a constant b, which obtains a low-loss space between the sub-model and the main model through regularization. Construct another constant c to avoid instability of the main model when the sub-model is updated.

[0076] The total loss function of the model is obtained as follows:

[0077]

[0078] L(h(x),y,W main )+bcos 2 (W main , W sub )+c‖W main -W * ‖ 2

[0079] Where h(x) is the hypothesis function of the model, which is used to calculate the loss between the model output y, and x is the model input. main and W sub are the model parameters of the main model and sub-model respectively, W * As the optimal parameters of the main model, the model parameters can be solved and gradually optimized in the iterative process.

[0080] The model training is implemented by gradually optimizing and iterating the loss function. On this basis, each sub-model is updated by mixing the main model and the sub-model. The update method is:

[0081] W sub =(1-a)W sub +aW sub

[0082] As mentioned above, a is the mixing weight, and the sub-models are mixed in turn during each iteration.

[0083] Alternatively, each master model performs updates based on the parameters of the sub-models.

[0084] In each update process, the mixed weight parameters are updated step by step. The way the mixed weight is updated is related to the local loss function of the sub-model.

[0085] In summary, the mixing weight of the mixing layer is:

[0086] W=ηW main +(1-η)W sub

[0087] Disparity information is highly correlated with the local features of a specific scene, while RGB images often contain more global context information, which helps to understand the scene as a whole. This mixed weight combination of local information and global information can effectively balance local details and global consistency;

[0088] In one embodiment, an attention mechanism is used as an adaptive module to generate the weight of each feature. The adaptive module dynamically generates weights based on the distribution characteristics and contextual information of the input features. The generated weights are used to perform weighted operations on the extracted features. In the sub-model, the features of different encoders are fused together through a smooth weighting strategy so that they complement each other in space. During model training, the weight adjustment module is continuously optimized to reduce the loss of information during feature fusion. Through the feedback of the loss function and the optimization algorithm, the weight of each feature is gradually adjusted so that the model can minimize the loss in the joint feature space.

[0089] Preferably, the output of multiple disparity transformation modules and the image to be tested are used to implement non-sensing detection of goods in the power parts warehouse, including: obtaining a target recognition network model based on an encoder-decoder architecture; the encoder is responsible for extracting semantic information and contextual features from the input image, and the decoder maps these features back to the original resolution to generate the final segmentation result.

[0090] The present invention uses bounding box detection. The position information of the bounding box Lloc (Lloc∈R4), i.e., the geometric center coordinates (cx, cy) and the length (w) and height (h) encoded as the four-node values ​​of the network, are predicted using a regression method. For each predicted bounding box, it is necessary to determine which category it contains. As with predicting bounding boxes, a set (c+1) of network nodes is used to predict the probability that the bounding frame belongs to each category (c+1 represents the number of data sets and background in the data set).

[0091] It can randomize the inspection data objects of the warehouse objects and the corresponding label data XML files before data preprocessing. During the neural network training process, random gradient descent is required to update the parameters of the network, and the randomization of image randomization can improve the randomness of the gradient descent direction. Thereby improving the convergence speed of the network and reducing the training time of the network. Before calculating the mean, the data set is divided into three parts, one is the training set, the number of pictures accounts for 60% of the total, and the other two parts are the validation set and the test set, each accounting for 20% of the total number of pictures. Then calculate the mean of the training set. When training the network, first subtract the image of the network from the mean, and then train and test. It also helps the speed and performance of network training.

[0092] In one embodiment, 4620 images of actual logistics warehouses were collected using the open source data annotation tool labelImg, which can be directly used to create XML files from rectangular boxes. It is similar to the form of the Pascal VOC 2012 dataset. The collected objects, such as the main points of study and pallets, are displayed in rectangular boxes. The corresponding labeled XML files are also generated. Finally, the collected warehouse photos are labeled to form a warehouse dataset for training and testing pallet detection algorithms.

[0093] Before data preprocessing, the inspection data objects of warehouse objects and the corresponding label data XML files can be randomized. In the process of neural network training, random gradient descent is required to update the parameters of the network, and the randomization of image randomization can improve the randomness of the gradient descent direction. This improves the convergence speed of the network and reduces the training time of the network. Figure 2 The improved convolutional neural network structure in the paper is used for parameter training. During the training process, the setting of the learning rate is very important. If the learning rate is too large, the convergence speed is very fast, but the training error will fluctuate and cannot converge to the global optimal value. If the learning rate is too small, the network converges very slowly and it takes a long time to reach the optimal value. Generally speaking, during the entire training process, a strategy of dynamically adjusting the network learning rate can be adopted. When we start training, it can set a higher learning rate so that the network quickly reaches a lower optimization level, and then reduce the learning rate. So that the network reaches the intermediate optimization level at a medium convergence speed, and then further reduce the learning rate, so that the network gradually reaches the optimal state. According to the design principle of the vector, the vector training settings are as follows: (i) the initial number of iterations (0-100,000) is set to 0.01; (ii) the intermediate number of iterations (100,001-300,000) is set to 0.001; (iii) the final optimization stage (3000-480,000) is set to 0.0001.

[0094] In this case, it will use our own labeled pallet dataset, and the network is able to converge at a rate of more than 300,000 iterations, and as the number of iterations increases, the loss decreases rapidly and the test MAP increases rapidly. After the final optimization stage, the convergence rate gradually decreases. The rate of loss reduction and MAP increase is significantly reduced, and the loss and MAP tend to stabilize, making the parameters of the network tend to be stable and basically optimal and converge to a fixed value, with a MAP index of 67.4% on the pallet test set. In addition, it can use the same network we have been studying and train and test in the form of the PASCAL VOC2012 public dataset, and the MAP index can reach 76.9%. Obviously, the test results of the released dataset are better than the pallet dataset. Because the quality of our labeled dataset is not well set. Many smaller pallets appear in our pallet dataset, which increases the difficulty of detection and also reduces the test MAP index. The computer CPU used in the experiment is Intel Core i7-6700k, with a main frequency of 4.0GHz, and the GPU model used is TITAN X, with a development capacity of 12G. The training iterations were 480,000 times, taking 50 hours, and the test frame rate could reach 42FPS.

[0095] The second aspect of the present invention relates to a detection system for the entry and exit of goods in a contactless warehouse using the method described in the first aspect of the present invention; the system includes an acquisition module, an analysis module, a transformation module, and a detection module; wherein the acquisition module is used to collect image information of a power parts warehouse, and obtain a test image through image labeling, image preprocessing, image enhancement, etc.; the analysis module is used to analyze the test image through a polygonal approximation method to calculate the contour of the target to be tested; the transformation module is used to apply a neural network to extract a feature map, and use a coding and decoding method to identify the target to be tested, thereby obtaining multiple parallax transformation modules; the detection module is used to implement contactless detection of goods in the power parts warehouse through the output of multiple parallax transformation modules and the test image.

[0096] In the present invention, the system uses real-time image processing and object recognition technology to achieve fast and accurate cargo detection, tracking and positioning, thereby optimizing the internal logistics operation process of the warehouse. Automated in-and-out warehouse control and intelligent cargo management greatly improve the warehouse's operating efficiency and response speed. The system can update and manage the location, quantity and status information of the goods in real time, avoiding errors and delays that may occur in traditional manual records. For warehouses with diverse cargo types and sizes, accurate inventory management is the key to keeping the supply chain running smoothly. The computer vision-based system can automatically detect abnormal situations, such as cargo detention and damage, and promptly issue alarms and take necessary measures. This real-time monitoring function helps reduce losses and improve the safety and risk management capabilities of warehouses.

[0097] Automated operations reduce manual intervention and operating costs, and reduce the overall cost of logistics management through more efficient resource utilization and transportation planning. At the same time, the high efficiency and reliability of the system also bring stable economic benefits and return on investment in long-term operations.

[0098] The system provides detailed real-time feedback and data reports to help managers gain in-depth understanding of all aspects of warehouse operations. These data not only support daily decision-making, but also provide reliable data support for long-term strategic planning and efficiency improvement, making management decisions more accurate and effective.

[0099] The system continuously receives and analyzes real-time data to optimize the training and parameter adjustment of the convolutional model to improve the recognition rate and reduce misidentification. This continuous optimization process ensures that the system can adapt to the ever-changing warehouse environment and business needs, maintain efficient operation and improve long-term stability.

[0100] The model of the present invention uses convolution technology in deep learning to perform image processing and object recognition on real-time videos in the warehouse. Convolution can effectively capture and learn features in images through multi-level convolution and pooling operations, so that the system can efficiently identify and track various types of goods, equipment and personnel. The convolution model learns the visual features of specific objects from a large number of samples through training, so as to quickly and accurately detect and identify targets in real-time video streams. This real-time and accuracy is the key to optimizing warehouse operation efficiency and response speed, especially in fast-paced and highly dynamically changing warehouse environments. The convolution model can handle complex scenes where multiple objects appear in the video frame at the same time, ensuring that all objects are accurately detected and classified. At the same time, the model has a certain scene adaptability and can maintain stable recognition performance under different lighting conditions, viewing angles and background noise.

[0101] A third aspect of the present invention relates to a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0102] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect of the present invention.

[0103] 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 them. Although the present invention is described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention still include contents that can modify or replace the specific embodiments of the present invention. Any modification or replacement that does not depart from the spirit and scope of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for detecting goods entering and leaving a warehouse without a sensor, characterized in that: The method comprises the following steps: Collect image information from the power parts warehouse and obtain the image to be tested through image labeling, image preprocessing, image enhancement, etc. The image to be measured is analyzed by polygonal approximation method to calculate the contour of the target to be measured; A neural network is used to extract feature maps, and a coding and decoding method is used to identify the target to be measured, thereby obtaining multiple disparity transformation modules; Through the output of multiple parallax transformation modules and the images to be tested, non-contact detection of goods in the power parts warehouse is implemented.

2. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 1, characterized in that: The image information of the power parts warehouse is collected, and the image to be tested is obtained through image labeling, image preprocessing, image enhancement and other work, including: Manually label the power accessory images to obtain product names of the power accessory images; Image preprocessing includes determining the effective area of ​​the electronic image through grayscale transformation and binarization, and then drawing the minimum circumscribed rectangle of the effective area. By calculating the deflection angle of the rectangle, the original electronic image is corrected to solve the tilt and distortion. Image enhancement obtains the image rotation matrix through the getRotationMatrix2D function, and obtains the rotated image through the affine transformation function.

3. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 2, characterized in that: The method of analyzing the image to be measured by a polygonal approximation method to calculate the contour of the target to be measured includes: The image is filtered using Gaussian filtering method; The optimal threshold is used to divide the image grayscale histogram into two parts so that the variance between the two parts is maximized to separate the target from the background.

4. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 3, characterized in that: Create the chord of the target curve. Assume that the first point of the curve is a and the last point is b. Connect the first point with the last point to form a straight line AB, which is the chord of the target curve. Find the point C with the maximum distance between the target curve and the chord, so that the maximum distance between the curve and the chord is d; Compare d with a preset threshold. If d is less than the specified threshold, the straight line can be approximated as the target curve, and the processing of the curve is completed. If the distance d is greater than the threshold, the line segment is divided into two segments, AC and BC, and C is used as the point. The above operation is continued until all curves are processed; The polygon is fitted by Douglas-Peucker algorithm, and the polygon approximation algorithm is used to identify the shape of the object.

5. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 4, characterized in that: The neural network is applied to extract feature maps, and the encoding and decoding method is used to identify the target to be measured, thereby obtaining multiple parallax transformation modules, including: The neural network involves convolution, activation function, pooling operation and out module; Taking the two-dimensional image signal I as input, slide the box of the same size and convolution kernel from the upper left corner to the lower right corner of the input image; Each time you slide the selected image region and perform convolution, check the multiplication and addition results of the corresponding elements; Calculate the new value S(I,j) in the input activation function f(S(I,j)), and the activation function selects the Sigmoid function, the hyperbolic tangent function, or the rectified linear unit function; Reduce the input size of the next layer, reduce the amount of calculation and the number of parameters, and obtain maximum pooling; When training a neural network, some hidden layer nodes are randomly selected and ignored with a preset probability.

6. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 5, characterized in that: The neural network is applied to extract feature maps, and the encoding and decoding method is used to identify the target to be measured, thereby obtaining multiple parallax transformation modules, including: The convolution 10_2 in the original SSD network and the features obtained from the convolution 9_2 are removed; Use the basic network VGG16, add convolution 6, convolution 7, convolution 8_2 and convolution 9_2 in sequence; Four different scale features of VGG16, Convolution 7, Convolution 8_2, and Convolution 9_2 are obtained to predict the position of the target bounding box and the corresponding confidence level.

7. A method for detecting goods entering and leaving a warehouse without a sense of contact according to claim 6, characterized in that: The method of implementing non-sensing detection of goods in a power parts warehouse through the outputs of multiple parallax transformation modules and the images to be tested includes: Obtain an object recognition network model based on an encoder-decoder architecture; The encoder is responsible for extracting semantic information and contextual features from the input image, while the decoder maps these features back to the original resolution to generate the final segmentation result.

8. A detection system for goods entering and leaving a non-contact warehouse using the method described in any one of claims 1 to 7; characterized in that: The system includes a collection module, an analysis module, a transformation module, and a detection module; wherein, The acquisition module is used to collect image information of the power parts warehouse and obtain the image to be tested through image labeling, image preprocessing, image enhancement, etc. The analysis module is used to analyze the image to be measured by a polygonal approximation method to calculate the contour of the target to be measured; The transformation module is used to extract feature maps using a neural network and identify the target using a coding and decoding method, thereby obtaining multiple parallax transformation modules; The detection module is used to implement non-contact detection of goods in the power parts warehouse through the outputs of multiple parallax transformation modules and the images to be tested.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images