An express parcel detection system based on Faster RCNN and its working method

Abstract

The present invention discloses an express parcel detection system based on Faster RCNN and its working method, which relates to the technical field of artificial intelligence and express logistics. This system is for enterprise-level user terminals; the enterprise-level user terminal includes a login management module, a parcel module, a function module for displaying detection results, a function module for prompting picking information, and a brightness equalization detection module. The present invention adopts an improved Faster RCNN, combines the position detection and category detection of express parcels with parcel form information, can save the separate recognition of form information by sensors, and can complete sorting through a monocular camera. And aiming at the low-precision problem caused by environmental factors to express parcels, an exponential mean brightness equalization algorithm capable of automatically adjusting brightness is proposed, which not only improves the detection accuracy of the model, but also facilitates the staff to monitor the parcels.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and express logistics, and particularly relates to an express package detection system based on Faster RCNN and its working method. Background Art

[0002] With the sharp increase in the number of packages, manual sorting can no longer meet the high-speed development of the express industry, and an express sorting system has emerged as the times require. The sorting system based on forms includes automatic sorting, manual sorting, and semi-automatic sorting. Automatic sorting is detected by radio frequency identification infrared barcodes, which requires a lot of miscellaneous hardware, high costs, and is difficult to promote. Manual sorting is directly completed by labor, with low efficiency. Therefore, most small and medium-sized logistics companies will adopt the semi-automatic sorting method. The main process is mainly that workers place the waybill face up, the gantry reads the form information, the information is synchronized to the control system, and the sorting is completed by a pushing device or a gravity sorting device.

[0003] However, the existing logistics sorting systems have the following defects:

[0004] 1. Lack of a monitoring system for express package sorting;

[0005] 2. The pushing devices and gravity sorting devices for semi-automatic sorting have relatively high costs, and separate sensors are also required to read the form;

[0006] 3. The existing systems have poor scalability. Express sorting can only be carried out at transfer stations and cannot be extended to express package warehousing, etc. Summary of the Invention

[0007] In view of this, the present invention provides an express package detection system based on Faster RCNN and its working method, which combines the position detection, category detection of express packages with the package form information, can save the separate recognition of form information by sensors, and can complete sorting through a monocular camera.

[0008] To this end, the present invention provides the following technical solutions:

[0009] The present invention provides an express package detection system based on Faster RCNN, and the system is for enterprise-level user terminals; the enterprise-level user terminals include:

[0010] Parcel Detection Module: It is used to obtain the express parcel pictures collected by the camera from the front end, put them into the saved parcel detection model for prediction, and return the predicted result pictures to the front end at the back end and present the pictures before and after detection on the web page; the parcel detection model is based on the improved Faster RCNN to detect express parcels; the constructed detection dataset is used to train the parcel detection model multiple times, select the best model weights, and save the detection model and weights;

[0011] Form Recognition Module: Use the Hough detection algorithm to establish a form recognition model; obtain the express parcel pictures collected by the same camera from the front end, and use the form recognition model to recognize the form information of the parcel;

[0012] Prompt Picking Information Function Module: It is used to analyze the detection results of the parcel detection module and the form recognition module, and judge and prompt whether the express parcels in this field of view can be grabbed by the number of returned categories and form numbers; the detection results include: parcel detection result pictures, detection position information, detection category information, detection probability and form information.

[0013] Furthermore, the enterprise-level user terminal further includes: a brightness balance detection module; the brightness balance detection module includes a picture preprocessing method that can handle pictures in both dark and overexposed environments, and is used to preprocess the express parcel pictures collected by the camera.

[0014] Furthermore, the picture preprocessing method is the exponential mean brightness balance algorithm; the exponential mean brightness balance algorithm includes:

[0015] Calculate the brightness average value of all datasets, and scale the average value to between (0, 1) to obtain the scaled brightness average value;

[0016] Obtain the brightness value of a single picture, normalize the brightness value of the single picture to between (-1, 1) to obtain the normalized brightness value of the single picture;

[0017] Take the gain of the normalized brightness value of a single picture on the exponential function.

[0018] Furthermore, the enterprise-level user terminal further includes: a login management module for registering accounts, modifying passwords, deleting passwords and verifying captcha security.

[0019] Furthermore, the enterprise-level user terminal further includes: a display detection result function module for displaying the detection results.

[0020] Further, determine and prompt whether the express packages in this field of view can be grabbed based on the number of returned categories and the form number, including: if the number is greater than 1, give prompts according to the different numbers of packages; if none are detected, directly prompt that there are no packages to be detected; if the number is 1, prompt that the packages of this category can be normally picked.

[0021] Further, the improved Faster RCNN includes:

[0022] A feature extraction network for extracting features; the feature extraction network is a lightweight network MobileNet;

[0023] A region proposal network for generating candidate boxes based on the features extracted by the feature extraction network;

[0024] An ROI pooling for collecting the candidate boxes generated by the region proposal network and extracting the candidate boxes from the features extracted by the feature extraction network;

[0025] A classification and regression network for classifying the candidate box features extracted by the ROI pooling.

[0026] Further, the enterprise-level client also includes: a dataset construction module for constructing a detection dataset, and the detection dataset is the picture data of the packages on the conveyor belt collected at a fixed angle of the camera.

[0027] Further, the form recognition module is also used to improve the accuracy of form recognition by angle selection and interference cropping.

[0028] The present invention also provides a working method of the above-mentioned express package detection system based on Faster RCNN, including the following steps:

[0029] Establish a brightness equalization detection model based on the exponential mean brightness equalization algorithm;

[0030] Establish an express package detection model through the improved Faster RCNN detection algorithm;

[0031] Establish a form recognition model using the Hough detection algorithm;

[0032] Obtain the pictures of the packages on the conveyor belt collected by the camera, preprocess the pictures using the brightness equalization detection model, perform form recognition using the form recognition model, and judge whether to detect the packages through the obtained form information; if detection is to be performed, use the express package detection model for detection and return the detected results to the control system for grabbing; if the form information cannot be obtained, it will be prompted in the monitoring system that the form information cannot be recognized and grabbing cannot be performed.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention designs an express package dataset for specific scenarios and divides it into the package category with the largest number of existing categories, making the grasping speed more flexible during picking.

[0035] 2. Aiming at the low-precision problem of express packages caused by environmental factors, an exponential mean brightness equalization algorithm that can automatically adjust brightness is proposed, which not only improves the accuracy of model detection but also facilitates the staff to monitor the packages.

[0036] 3. By combining package detection and form recognition, the Faster RCNN object detection network is improved, and form recognition is also improved. It realizes low-cost equipment and high-precision detection, and can complete sorting through a monocular camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for description in the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. <G

[0038] Figure 1 It is a schematic structural diagram of an express package detection system based on Faster RCNN in an embodiment of the present invention;

[0039] Figure 2 It is a flowchart of the working method of an express package detection system based on Faster RCNN in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] In order to enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0042] The present invention provides a courier package detection system based on Faster RCNN, which is sorted by a gravity sorting device. First, on the basis of the original single-category package detection, the number of package categories is expanded, and the angle of picking and clamping can be flexibly selected. Secondly, in the same camera, form recognition is completed. The main application advantages include saving sensor overhead, flexibly selecting the clamping angle, and faster picking speed, etc. The main process can be simplified as follows: after the waybill is placed face up by the operator, the gravity device directly completes the sorting. The detection data set constructed by this system is to classify and detect the courier packages on the conveyor belt (mainly including 5 major categories, almost covering all existing courier types) by simulating at a fixed camera angle (the camera is vertically downward, about 1m away from the package height). The detection mainly obtains the category of the package, the category probability of the package, the position information of the package and the waybill number. According to the coordinate information of the package and the category, the grasping angle of the robotic arm clip is determined, and the package is conveyed to the control system according to the waybill information to obtain the position where the package is about to arrive. The present invention has completed the development of a courier package detection system based on Faster RCNN by using the Flask framework, the HTML+CSS front end and the Faster RCNN framework of PyTorch. This system is a website system based on Windows. This system mainly completes the detection of courier packages, including direct detection and brightness equalization detection. The system can display the detected pictures and the detection result information.

[0043] Flask is a lightweight, customizable framework written in Python. Compared to other similar frameworks, it's more flexible, lightweight, secure, and easy to use. It integrates well with the MVC model for development, allowing developers to work collaboratively and quickly. Small teams can quickly implement feature-rich small or medium-sized websites or web services. Flask is also highly customizable, allowing users to add features based on their needs, enriching and expanding functionality while keeping the core functionality simple. Its powerful plugin library allows users to customize their websites and develop powerful websites.

[0044] The key feature of the Flask framework is its relatively simple core structure, yet its strong scalability and compatibility allow programmers to quickly implement a website or web service using the Python language. Flask primarily includes two core libraries: Werkzeug and Jinja2, which handle business processing and security, respectively. These foundational functions provide a rich set of building blocks for web project development. The Werkzeug library is extremely powerful and comprehensive, supporting URL routing request integration, enabling it to respond to multiple user requests simultaneously. It also supports cookie and session management, establishing persistent connections through identity caching and improving user access speed. It also supports interactive JavaScript debugging, enhancing the user experience. It handles basic HTTP transactions and quickly responds to client-side requests. The Jinja2 library supports automatic HTML migration, effectively mitigating external script attacks. The system runs very quickly, and the page load process compiles source code into Python bytecode, enabling efficient template execution. Its template inheritance mechanism allows for modification and maintenance of template content, providing tailored templates for diverse user needs. Compared to other lightweight frameworks, the Flask framework offers excellent scalability, making it an unmatched choice.

[0045] Faster RCNN is a milestone network for target detection. The introduction of this network truly applies end-to-end deep learning methods to the field of target detection. It greatly improves the detection speed and accuracy of target detection. The system in this invention replaces the original VGG16 feature extraction network with a lightweight network MobileNet. For different feature extraction networks, comparative experiments are conducted on their own data sets many times. It can be found that MobileNet is not only highly accurate, but also much faster than other feature extraction networks. The Faster RCNN detection part can be mainly divided into four modules:

[0046] (1) backbone. That is, the feature extraction network, which is used to extract features. A set of conv+relu+pooling layers are used to extract the feature maps of the image for the subsequent RPN layer and the extraction of proposals. To improve the detection speed, MobileNetV2 is selected as the backbone of this system. MobileNetV2 is constructed using depthwise separable convolutions, which can reduce a large number of parameters of the model and reduce the latency of the model. It has been compared with the Resnet series many times. The research results show that although the mAP value of Resnet+FPN is 0.5 points higher than that of MobileNet, the speed of MobileNet is 10 times faster than that of ResNet.

[0047] (2) RPN (Region Proposal Network). That is, the region candidate network, which replaces the Selective Search in the previous RCNN version and is used to generate candidate boxes. There are two parts to this task. One is classification: judging whether all preset anchors belong to positive or negative (that is, whether there is an object in the anchor, binary classification); the other is bounding box regression: correcting the anchors to obtain more accurate proposals. Therefore, the RPN network is equivalent to doing part of the detection in advance, that is, judging whether there is an object (the specific category is not judged here), and correcting the anchor to make the box more accurate.

[0048] (3) RoI Pooling. That is, the region of interest pooling (spatial pyramid pooling in the SPP net), which is used to collect the proposals generated by the RPN (the coordinates of each box) and extract them from the feature maps in (1) (cropped from the corresponding positions), generating the proposals feature maps and sending them to the subsequent fully connected layers for further classification (specifically which category) and regression.

[0049] (4) Classification and Regression. Calculate the specific category using the proposals feature maps, and at the same time do another bounding box regression to obtain the final accurate position of the detection box. Although this framework was proposed a long time ago, its accuracy has always been at the forefront of various detection models. It is one of the most important object detection models, with high accuracy, strong scalability, and a large space for model optimization. Therefore, this model is selected for object detection.

[0050] Such as Figure 1As shown in the figure, an embodiment of the present invention provides an express package detection system based on Faster RCNN, mainly targeting enterprise-level user terminals. The enterprise-level user terminal includes:

[0051] Login management module 100: registering accounts, modifying passwords, deleting passwords, verifying security codes, etc.

[0052] Dataset construction module 200: constructing a detection dataset and storing it in a local file. The detection dataset is picture data of packages on a conveyor belt collected at a fixed angle of the camera.

[0053] Package detection module 300: obtaining the picture of the express package collected by the camera from the front end, putting it into the saved model for prediction, and returning the predicted result picture to the front end through the back end, and presenting the pictures before and after detection on the web page. The detection of express packages is realized by the package detection model based on Faster RCNN. The detection dataset constructed by the dataset construction module 200 is used to train the package detection model multiple times, the best model weights are selected, and the detection model and weights are saved.

[0054] Form recognition module 400: using the Hough detection algorithm to establish a form recognition model; obtaining the picture of the express package collected by the same camera from the front end, and using the form recognition model to recognize the form information of the package. Further, the accuracy of form recognition is improved by angle selection and interference object cropping.

[0055] Display detection result function module 500: The detection results mainly include: the package detection result picture, detection position information, detection category information, detection probability, form information, etc. After these detection results are obtained through the model, they are returned to the front end and displayed side by side.

[0056] Prompt picking information function module 600: analyzing the results after model detection, and judging whether the express package in this field of view can be grabbed based on information such as the number of returned categories and form numbers. If the number is greater than 1, the robotic arm may cause missed detection during the picking process, so prompts are given according to the different numbers of packages. If nothing is detected, it directly prompts that there are no detectable packages. If it is normally 1, it prompts that the package of this category can be picked normally.

[0057] Brightness Equalization Detection Module 700: It includes a picture preprocessing method for simultaneously processing pictures in dark and exposure environments, and is used to preprocess the express package pictures in the local files obtained from the front end. The brightness equalization algorithm has two main functions in this system. The first is that when using this system, it can observe the pictures to be detected more clearly and clearly, which not only reduces the harm to eyesight but also makes it easier to observe the pictures before and after detection. The second is that the input processed pictures can improve the accuracy of express package detection.

[0058] As Figure 2 shown, in the embodiment of the present invention, a working method of the above-mentioned express package detection system based on Faster RCNN is also provided. First, an express package detection model is established through an improved Faster RCNN detection algorithm; an adaptive brightness equalization algorithm is proposed and a brightness equalization module is added; then Hough detection is used to establish a form detection model, and the accuracy of form recognition is improved by angle selection and interference object cropping. Finally, it can detect forms and packages simultaneously. The specific steps are as follows:

[0059] S1. Establish an express package detection model through an improved Faster RCNN detection algorithm:

[0060] It is necessary to upload files and then perform detection to complete the detection. Modify the backbone network of the Faster RCNN network, which is specifically used to implement the detection of express packages. Train multiple times, select the best model weights, and save the detection model and weights. Obtain the express package pictures in the local files from the front end, put them into the saved model for prediction, and the resulting predicted pictures are returned to the front end at the back end and presented on the web page as the pictures before and after detection. Some vision-guided robot picking systems are deployed on edge devices with limited computing resources. It is crucial to effectively control the model parameters, computational volume, memory access volume, etc. of object detection at the edge with limited computing resources.

[0061] In order to improve the speed and accuracy of the model and reduce the number of model parameters, the present invention studies the lightweighting of the model. Lightweight networks mainly include knowledge distillation, weight quantization, pruning operation-based compressed models, and directly training lightweight network models. The present invention directly designs a lightweight network through depthwise separable convolution for the detection of express packages. In order to further improve the speed of the model, the anchor generation mechanism in the RPN network structure of Faster RCNN is modified. The main modification steps are as follows:

[0062] S11. Statistically count the frequency of the area and aspect ratio of the real annotation boxes;

[0063] S12. Select the interval value with the highest frequency of occurrence for the aspect ratio, and select the median and upper quartile for the wrapping area and scale it according to the input ratio. These two values serve as the main basis for the anchor mechanism;

[0064] S13. Finally, determine the anchor sizes for four combinations with aspect ratios of 1:1.2 and 1:1.6 and areas of 256 and 512;

[0065] S14. Change RoI Pooling to RoI Align.

[0066] S2. Establish a brightness balance detection model.

[0067] To address the impact of environmental factors during the sorting of express parcels and enable administrators to observe parcels more clearly when monitoring the sorting process, the present invention proposes an adaptive brightness balance algorithm. This not only improves the accuracy of express parcel detection but also facilitates staff monitoring.

[0068] In computer vision, image brightness processing mainly involves controlling the brightness values of pixels to perform gain or attenuation operations. Traditional brightness processing not only requires manual control but also cannot handle both over-bright and over-dark images simultaneously. Adaptive image brightness processing and histogram equalization can effectively solve this problem, but the disadvantages of these algorithms are that they can amplify image noise, increase the contrast of background interference information, and reduce the contrast of useful signals. To overcome the disadvantages of the above algorithms, in the embodiments of the present invention, an exponential mean brightness balance algorithm is proposed based on the traditional brightness gain method. The main steps to implement this algorithm include:

[0069] S21. To ensure the brightness of the image is constant and does not cause excessive interference to the detection, calculate the average brightness of all datasets and scale the mean to the range (0, 1) according to formula (1) to obtain x_avg. Here, b_min is the minimum pixel value, b_max is the maximum pixel value, and b_avg is the average pixel value.

[0070] S22. Obtain the brightness value of a single image and normalize it to the range (-1, 1) according to formula (2), where the brightness value is x.

[0071] S23. Take the gain of x in the range (-1, 1) on the exponential function, and the gain value is represented by h.

[0072]

[0073]

[0074] h = x_avg x (3)

[0075] S3. Use the Hough detection algorithm to establish a form recognition model;

[0076] In the embodiments of the present invention, the edge information of the form of the express package is extracted through the Canny algorithm. The goal of Canny is to find an optimal edge detection algorithm. The meaning of optimal edge detection includes: optimal detection: the algorithm can identify as many actual edges in the image as possible, and the probabilities of missing real edges and misdetecting non-edges are both as small as possible. Optimal positioning criterion: the position of the detected edge point is closest to the position of the actual edge point, or the degree to which the detected edge deviates from the true edge of the object due to noise influence is the smallest. One-to-one correspondence between detection points and edge points: the edge points detected by the operator should correspond one-to-one to the actual edge points.

[0077] In the embodiments of the present invention, the Hough transform is used to detect the straight-line position in the form edge information, obtain the inclination angle of the straight line, and then make the straight line horizontal clockwise or counterclockwise. Then the pyzbar library is called to obtain the form information of the package. The main formula for the Hough line transformation is as follows:

[0078]

[0079] r = x cosθ + y sinθ (5)

[0080] θ1 = arctant (6)

[0081] Where (x, y) are a series of discrete points. By using the voting method according to formula (5), the values of θ and r can be determined, thereby determining the slope of the entire straight line. Finally, according to formula (6), the rotation angle θ1 to be rotated is obtained, where t is the slope of the straight line.

[0082] Furthermore, the accuracy of form recognition is improved by angle selection and interference object cropping. Analyzing the results after the failure of Hough correction, it is found that the failure is because the straight line obtained by the Hough transform is not the straight line representing the edge of the form. Subsequently, dilation and erosion operations are performed on the picture to expand the range of the form, thereby determining the angle. The accuracy is improved, but there are still a small number of forms that cannot be recognized. Visualizing the process of the picture reveals that there are obvious line interferences at the background corners in the dataset. Crop the background picture.

[0083] S4. Combination of form recognition and package detection:

[0084] In semi-automatic sorting based on machine vision during parcel sorting, specific sensors are relied on to read form information and synchronize it to the control system to complete one sorting operation. To reduce hardware costs, in the embodiments of the present invention, express parcel detection and form recognition are carried out in one camera. Based on the obtained form information, it is determined whether to detect the parcel. If detection is to be carried out, the detected result is returned to the control system for grasping. If the form information cannot be obtained, a prompt will be given in the monitoring system that the form information cannot be recognized and grasping cannot be carried out. Parcels for which the form cannot be recognized will complete the next detection during the process design stage.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A courier package detection system based on Faster RCNN, characterized in that, The system is targeted at enterprise-level user terminals; the enterprise-level user terminals include: The package detection module is used to obtain express package images captured by the camera from the front-end, feed them into a saved package detection model for prediction, and return the predicted result images to the front-end through the back-end, with the pre- and post-detection photos displayed on the webpage. The package detection model detects express packages based on an improved Faster RCNN. The package detection model is trained multiple times using the constructed detection dataset, and the model weights are selected and saved. Form recognition module: Use the Hough detection algorithm to build a form recognition model; obtain express package images captured by the same camera from the front end, and use the form recognition model to identify the package form information; The picking information prompt function module is used to analyze the detection results of the package detection module and the form recognition module, and judge and prompt whether the express package in the camera's field of view can be picked up based on the number of categories and form numbers returned; the detection results include: package detection result image, detection location information, detection category information, detection probability and form information; The enterprise-level user terminal also includes: a brightness balance detection module; the brightness balance detection module includes an image preprocessing method capable of processing both dark and exposed environments, for preprocessing express package images captured by the camera; The image of the package on the conveyor belt captured by the camera is obtained, the image is preprocessed using the brightness balance detection model, the form recognition model is used to perform form recognition, and the obtained form information is used to determine whether the package is to be inspected; if inspection is to be performed, the express package inspection model is used to perform the inspection, and the inspection result is returned to the control system for capture; if the form information cannot be obtained, a prompt will be displayed in the monitoring system, indicating that the form information cannot be recognized and capture cannot be performed.

2. The express package detection system based on Faster RCNN according to claim 1, characterized in that, The image preprocessing method is an exponential mean brightness equalization algorithm; the exponential mean brightness equalization algorithm includes: Calculate the average brightness of all data sets and scale the mean to (0, 1) to obtain the scaled average brightness; Obtain the brightness value of a single image, normalize the brightness value of a single image to between (-1, 1), and obtain the normalized brightness value of the single image; For the normalized brightness value of a single image, gain is taken on the exponential function.

3. A courier package detection system based on Faster RCNN according to claim 1, characterized in that, The enterprise-level user terminal also includes a login management module for registering an account, changing a password, deleting a password, and performing verification code security verification.

4. The express parcel detection system based on Faster RCNN according to claim 1, characterized in that, The enterprise-level user terminal further includes: a detection result display function module for displaying the detection result.

5. The express parcel detection system based on Faster RCNN according to claim 1, characterized in that, The number of returned categories and the order number are used to determine and prompt whether the express parcels in the camera's field of view can be picked up, including: if the number is greater than 1, prompts are given based on the number of packages; if no detectable packages are found, a direct prompt is given that there are no detectable packages; if the number is 1, a prompt is given that the parcels of this category can be picked up normally.

6. The express parcel detection system based on Faster RCNN according to claim 1, characterized in that, The improved Faster RCNN includes: A feature extraction network for extracting features; the feature extraction network is a lightweight network MobileNet; Region Proposal Network that generates candidate boxes based on the features extracted by the feature extraction network; Region of Interest Pooling that is used to collect the candidate boxes generated by the Region Proposal Network and extract the candidate boxes from the features extracted by the feature extraction network; Classification and regression network that is used to classify the candidate box features extracted by the Region of Interest Pooling.

7. A courier package detection system based on Faster RCNN according to claim 1, characterized in that, The enterprise-level client also includes: a dataset construction module for constructing a detection dataset, which is picture data of packages on a conveyor belt collected by simulating a fixed angle of a camera.

8. The express parcel detection system based on Faster RCNN according to claim 1, wherein The form recognition module is also used to improve the accuracy of form recognition by angle selection and interference cropping.

Images