CN112115043A - Image-based on-end intelligent page quality inspection method - Google Patents

Abstract

The invention discloses an image-based on-end intelligent page quality inspection method, which comprises the steps of firstly obtaining DOM element rectangular frame coordinates, screenshots and rectangular frame coordinates of abnormal region positions of normal and abnormal pages, training a page abnormal problem classification and detection model and an element semantic recognition model, then obtaining all DOM element position information and screenshots of a page to be inspected, presorting various category modules according to DOM element distribution characteristics, adopting the element semantic recognition model to recognize a universal module, adopting the trained detection model to directly detect empty pit problems and file overlapping problems, recognizing empty screen problems according to effective SURF characteristics, identifying empty floor problems through the distance of adjacent floor titles, and calculating the similarity of commodity resource positions to judge the same element problems. The invention can liberate testing personnel to carry out repeated manual testing work, and has remarkable effects of reducing cost and improving efficiency on testing work and on-line monitoring and inspection work.

Description

An image-based on-device intelligent page quality inspection method

technical field

The invention belongs to the fields of testing technology and computer vision technology, and in particular relates to an image-based on-device intelligent page quality inspection method.

Background technique

With the vigorous development of the mobile Internet, the commodity object information service system has shifted from the PC era to the wireless era. In the digital e-commerce industry, due to the wide variety of marketing promotion activities and flexible changes, the H5 page release of the venue of the marketing activities often supports the marketing rule makers (non-technical personnel) to flexibly create or change the configuration at any time according to operational and marketing needs. The flexibility of change may bring greater online quality risks, and the online quality of these pages often directly affects user experience, marketing campaign effectiveness, and even sales. However, the strong decentralization of the page greatly increases the difficulty of quality assurance, especially during the promotion period, hundreds or even thousands of H5 events/site pages are changed and released thousands of times, resulting in extremely high cost of human testing (multi-person interactive communication, multi-platform switching) operations, tasks are fragmented and dispersed). Therefore, a general and low-cost solution is needed to effectively detect common serious problems such as empty screens, empty windows (missing some resource bits), and "empty floors" (missing module content) that may appear on the page, so as to efficiently ensure the online quality of the page. .

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an image-based on-device intelligent page quality inspection method in view of the deficiencies of the prior art.

The object of the present invention is achieved through the following technical solutions: an image-based on-end intelligent page quality inspection method, comprising the following steps:

S1: Obtain DOM element location information and page screenshots of the page, where the page screenshots include normal page screenshots and abnormal state page screenshots; the abnormal state screenshots include page screenshots containing empty pits and page screenshots containing copywriting overlay problems; And additionally obtain the DOM element position information corresponding to the empty pit area and the overlapping coverage area of the copy;

S2: According to the DOM element location information and page screenshots obtained in step S1, respectively train a strip module classification model, a resource bit module classification model, an anomaly detection model for empty pits, and a copywriting overlap coverage recognition model;

The training process of the strip module classification model and the resource bit module classification model is specifically: according to the DOM element position information and page screenshots obtained in step S1, screening and intercepting to obtain strip elements and resource bit elements; The corresponding category is trained by deep learning residual network or automatic machine learning technology to obtain a strip module classification model, the input is strip element, and the output is strip element category; according to resource bit elements and their corresponding categories, deep learning residuals are used The network or automatic machine learning technology is trained to obtain the resource bit module classification model, the input is the resource bit element, and the output is the resource bit element category;

The categories of the strip elements include elevator navigation module, floor title, bottom Tab navigation bar and other categories; the categories of the resource element include commodity resource category, coupon category, store category and other categories;

The training process of the empty pit anomaly detection model is specifically as follows: after executing S1 in batches, firstly, a round of manual review is performed on the screenshot of the page containing the empty pit and the label of the bounding box coordinates of the DOM corresponding to the empty pit element obtained in S1. That is, delete the samples that are actually non-empty pit labels and add the unlabeled empty pit areas, and obtain the reviewed page screenshots containing empty pits and their marked bounding box sample data, and use the R-FCN target detection framework to train to get Empty pit detection model, the input of the model is an image, and the output is whether it contains an empty pit;

The training process of the copywriting overlapping recognition model is specifically as follows: after obtaining the screenshot of the page containing the copywriting overlapping coverage problem and the DOM element position information of the problem area obtained in step S1, using the R-FCN target detection framework to train to obtain the copywriting overlapping coverage class detection Model, the input is a screenshot of the page, and the output is whether there is overlap of copywriting;

S3: Obtain the position information of the DOM element of the page to be checked and the screenshot of the page to be checked, and filter and intercept the strip element to be checked and the resource bit element to be checked;

S4: Using the strip module classification model and the resource bit module classification model obtained in step S2, respectively classify the strip elements to be checked and the resource bit elements to be checked obtained in step S3;

S5: Identify problems on different pages, specifically:

(1) for the empty pit class page problem, utilize the empty pit class anomaly detection model obtained in step S2 to detect whether there is an empty pit on the page to be checked obtained in step S3;

(2) For the problem of empty floor pages, determine whether the number of floor title bar elements classified in step S4 is greater than or equal to 2 and the longitudinal relative distance between two adjacent floor title bar elements is less than the set empty floor threshold ; If yes, there is an empty floor on the page to be checked, otherwise there is no empty floor;

(3) For the blank screen type page problem, take the middle area of the screenshot of the page to be checked obtained in step S3, and extract the SURF feature key points in the middle area. If the number of SURF feature key points extracted is less than the set key point If the threshold is set, it is judged that there is a blank screen problem on the page to be checked, otherwise there is no blank screen problem;

(4) For the problem of duplicate materials, compare the similarity of every two commodity resource bit class resource bit elements classified in step S5, and determine whether there is an abnormal problem that different commodity resource bit class resource bit elements are associated with the same target object information, and the similarity If it is greater than the set threshold, it is abnormal;

(5) For the problem of overlapping pages of copywriting, use the overlapping and covering recognition model obtained in step S2 to check whether the screenshot of the page to be checked obtained in step S3 has overlapping coverage of copying.

Further, use the strip module classification model and the resource bit module classification model to identify strip elements and resource bit elements of different categories in the page, and set the corresponding interactive control positions; configure the interactive control positions and sequences that need to be checked, according to the configuration. Simulate clicking the specified interactive control position on the page and obtain the DOM element position information and page screenshot of the clicked page; if the specified clicked interactive control position does not exist in the page screenshot of the current screen, perform a sliding operation to reach the next screen position , continue to simulate clicking the specified interactive control position and obtain the DOM element position information and page screenshot of the clicked page.

Further, using the strip module classification model and the resource bit module classification model to identify strip elements and resource bit elements of different categories in the page, and setting the corresponding interactive control positions are specifically:

(a) The position of the identified resource position element of commodity resource class, coupon class and store class is directly used as the position of its interactive control;

(b) For the identified strip elements of the elevator navigation module, use the OCR technology to identify the position of the interactive controls in the elevator navigation module for reaching a certain floor, and use the template matching method to identify all the contents used to expand the elevator navigation module. The interactive control position of the arrow button of ;

(c) Using OCR technology to identify the identified bottom Tab navigation bar strip elements to obtain the interactive control positions for jumping to other column pages.

Further, the template matching method is SURF feature extraction algorithm combined with FLANN.

Further, the DOM element position information includes the relative coordinates of the rectangle bounding box of the DOM element in the page screenshot, the aspect ratio of the rectangle bounding box, the ratio of the sides of the rectangle bounding box to the screenshot, and the area of the rectangle bounding box.

Further, the screening condition of the strip element is w/h∈(5,6), w/W∈[1,3); the screening condition of the resource bit element is h/w<3 and w/h <3.5 and 15W<h*w<150W; w is the width of the element, h is the height of the element, and W is the width of the image where the element is located.

Further, the threshold value of the empty floor in the step S5 is 10-20 pixels.

Further, the height and width of the middle area in the step S5 are both greater than half of the height and width corresponding to the screenshot of the page to be checked.

Further, the threshold value of the key point in the step S5 is 10-100.

Further, in the step S5, the similarity of every two commodity resource bit class resource bit elements classified in the comparison step S4 is specifically: extracting the HOG feature of the commodity resource bit class resource bit element classified in step S4, calculating The cosine similarity of the HOG features of each two commodity resource bit class resource bit elements.

The beneficial effects of the present invention are as follows: the present invention is a universal automatic inspection solution for common page quality problems of e-commerce event venue pages, and is generally applicable to both offline store-type emerging businesses and traditional e-commerce businesses. The ability to check common page abnormal problems is to simulate the visual inspection ability of testers for abnormal problems; to detect and identify the position of interactive controls, and to click on the controls through JavaScript scripts, which can realize position switching and floating layers in the page. View changes such as pop-ups and automatic jumps between pages are liberating the operation cost of testers. Therefore, the present invention realizes automatic online page quality inspection. During the period when e-commerce activities are concentrated, for large-scale online page changes, using the present invention can free testers to perform repetitive manual testing work, which will affect the testing work and Online monitoring and inspection work has a significant effect of reducing costs and improving efficiency.

Description of drawings

1 is a schematic flowchart of an image-based on-device intelligent page quality inspection method according to the present invention;

FIG. 2 is a schematic diagram of an application example of an image-based on-device intelligent page quality inspection method of the present invention;

Fig. 3 is a schematic diagram of abnormal problem identification of this method; wherein, (a) is a schematic diagram of an empty pit problem; (b) is a schematic diagram of an empty screen problem; (c) is a schematic diagram of an empty floor problem; (d) is the same material problem Schematic.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art fall within the protection scope of this application.

The present invention is an image-based on-device intelligent page quality inspection method, which includes several learning-based image classification models and image target detection models to classify common modules of pages and identify abnormal states of pages; wherein, the classification of strip modules The model, resource location module classification model, and copywriting overlap/coverage problem recognition model are image classification models, and empty pit anomaly detection models are image target detection models. The above learning-based models include a training phase and a testing phase (ie, the application phase), such as As shown in Figure 1, it includes the following steps:

S1: Run a script on the mobile device to save the page's DOM element location information and screenshots.

In the training phase of the learning-based model, the position information and screenshots of all DOM elements within the current screen display range of the H5 page are saved by running the JavaScript script on the mobile device, and the page sliding is controlled; the DOM element position information includes the rectangle of the DOM element. The relative coordinates of the bounding box in the screenshot, the aspect ratio of the rectangular bounding box, the ratio of the length/width of the rectangular bounding box to the length/width of the corresponding screenshot, and the area of the rectangular bounding box; the screenshots include normal pages. Screenshots and screenshots of abnormal pages formed by manipulating and modifying the HTML structure and CSS of page elements. The abnormal state screenshots include screenshots of pages containing empty pits and images containing overlapping coverage of copywriting. The JavaScript script additionally saves the coordinates of the DOM bounding box corresponding to the empty pit element and the sample data of its marked bounding box, as well as the overlapping coverage of copywriting. The coordinates of the bounding box of the DOM corresponding to the region and the sample data of its marked bounding box. The empty pit means that the content of the resource slot is empty or lacks a resource slot; for example, the content of the resource slot is empty if the main information such as the product displayed on the page and the pictures in the area of the store is missing; the modules used in a certain area are displayed in three rows in a row. The specification module of 1 resource bits actually only shows that two resource bits belong to the missing resource bits.

S2: After marking or precipitating the image data in the normal state and abnormal state of the page, train the strip module classification model, the resource bit module classification model, the empty pit type anomaly detection model and the copywriting overlap coverage recognition model respectively;

The training process of the strip module classification model and the resource bit module classification model is specifically:

(1) For the training of the strip module classification model and the resource bit module classification model, use the screenshot output by S1 and the position information of all DOM elements within the screenshot range, and screen out the strip element ROI (Region) that meets the conditions according to certain rule constraints of Interest, region of interest) and resource bit element ROI. The filter condition of the strip element is w/h∈(5,6), w/W∈[1,3); the filter condition of the resource bit element is h/w<3 and w/h<3.5 and 15W<h*w<150W; w is the width of the element, h is the height of the element, and W is the width of the image where the element is located.

(2) After obtaining the strip element ROI in (1) in batches, mark it according to its element type. Common strip element categories include elevator navigation module, floor title, bottom Tab navigation bar and other categories; similarly, for resource Bit element ROI is used for marking. Common resource bit element categories include commodity resource bits, cards and coupons, store categories and other categories. The elevator navigation module is a navigation module for realizing that clicking on the corresponding category will jump to the following specific floor; the other categories are divided according to the common page layout modes and common modules of different services.

(3) After obtaining the batch marking data described in step (2), a mature deep learning residual network Resnet or automatic machine learning AutoML technology is used to train the strip module classification model and the resource position module classification model respectively. The input of the strip module classification model is the strip element, and the output is the strip element category; the input of the resource bit module classification model is the resource bit element, and the output is the resource bit element category.

The empty pit type anomaly detection model is used to check whether there is a problem of empty pits or dropped pits on the page. The training process is specifically as follows: after executing S1 in batches, first, the screenshots of the pages containing empty pits obtained in S1 and the corresponding empty pit elements are performed. A round of manual review and marking is performed on the labels of the bounding box coordinates of the DOM, that is, the samples that are actually non-empty pit labels are deleted and the unlabeled empty pit areas are added to obtain the reviewed page screenshots containing empty pits and their marked bounding boxes. After the sample data, the R-FCN target detection framework is used to train the empty pit detection model. The input of the model is an image, and the output is whether it contains empty pits.

The specific training process of the copywriting overlap recognition model is as follows: after executing S1 in batches, firstly perform a round of manual review and marking on the bounding box coordinate label of the area containing the copywriting overlapping display problem obtained in S1, that is, delete the actual non-copywriting Overlapping and covering samples and increasing the overlapping coverage area of labeled and unlabeled texts, after obtaining the reviewed images with copywriting overlapping coverage problems and their labeled bounding box sample data, the R-FCN target detection framework is used to train to obtain a copywriting overlapping coverage class detection model , the input is an image, and the output is whether there is a copy overlay.

S3: After the training of each model is completed, in the application stage, run the JavaScript script on the mobile device to save the coordinate information and screenshots of the bounding boxes of all DOM elements within the current screen display range in the H5 page to be checked;

S4: Pre-classify each category module according to the position distribution characteristics of DOM elements, and for the image output by S3 and the bounding box coordinate information of all DOM elements within the image range, according to the aspect ratio and area of each bounding box, the size of the bounding box in the image The relative position and other features are used to screen out the candidate strip element ROI and the candidate resource bit element ROI.

S5: Using the strip module classification model and the resource bit module classification model obtained in S2, classify the strip elements and resource bit elements obtained in S4 respectively, that is, common element semantic recognition.

S6: Identify problems on different pages, specifically:

(1) For the problem of empty pit pages, use the empty pit abnormal detection model obtained in S2 to detect the current screen shot obtained in S3 to determine whether there are empty pits.

(2) For the problem of empty floor pages, if the number of floor title strip elements output by S5 is greater than or equal to 2, and the longitudinal relative distance Δy of the ROI of two adjacent floor title strip elements is less than a certain threshold, usually 10 ～20 pixels, it is judged that there is an "empty floor" on the page, which means that the content of the entire resource bit module is empty.

(3) For the problem of blank screen pages, take the ROI in the middle area of the current screenshot in S3 in space, and the height and width of the ROI in the middle area are greater than 1/2 of the height and width of the current screenshot in S3; in the middle area In the ROI, extract the SURF feature key points. If the number of SURF feature key points extracted is less than the set threshold, the threshold is set according to the actual middle area ROI scale and specific business demand scenarios, usually 10 to 100, and the page is judged There is a blank screen problem.

(4) For the problem of duplicate materials, take all the commodity resource bit class element ROIs output by S5, compare the similarity of commodity resource bit class element ROI in different regions, extract the HOG feature of each commodity resource bit class element ROI, and calculate all the commodity resource bit class elements in pairs. The cosine similarity of the HOG feature of the commodity resource bit class element ROI is used as the similarity calculation result of each pair of commodity resource bit class element ROI, and the similarity threshold is set to determine whether there is an abnormal problem that different commodity resource bit elements are associated with the same target object information. Similarity greater than the threshold is abnormal.

(5) Aiming at the problem of overlapping and covering pages of copywriting, the current screen shot obtained in S3 is detected by using the overlapping and covering recognition model of copying obtained in S2 to determine whether there is overlapping and covering of copying.

If it is identified in the above S6 that the page has an abnormal problem as shown in FIG. 3 , the page quality alarm information is output. At the same time, the user can configure the specified operation path in the system for the system to automatically check the page. If you want to click any store type control from the current event page of a certain venue to jump to the store acceptance page to continue the inspection, you need to check in the system. Configure "shop class" as a specified class for the control that the script will perform the click action on next.

For the position of the next operation interaction to be performed by the system JavaScript script, identify the interaction operation points by type, and use the element semantic recognition model to identify areas such as commodity resource slots, coupons, store resource slots, etc., which are directly used as the interaction operation points of this category; The element semantic recognition model identifies the ROI of the elevator and the bottom Tab navigation bar, and then combines the OCR technology and template matching method as a module to click directly to the interaction point and other main columns to click directly to the interaction point, specifically:

(1) The elevator navigation module is obtained by identifying the S5 strip module classification model. In this module, the OCR optical character recognition technology is used to detect and identify the controls of a certain floor in the elevator navigation module; the template matching method is used to identify and expand the elevator navigation module. Arrow buttons for all content, such as SURF (Speeded Up Robust Features) feature extraction algorithm combined with Fast Approximate Nearest Neighbor Search Library (FLANN).

(2) Bottom Tab navigation module, the bottom Tab navigation bar module identified by the S5 strip module classification model, OCR technology is used in this module, and by capturing the position of key text, some interactive actions can be obtained after clicking. The position of the control is the position of the interactive control. The interactive action generally refers to jumping to other column pages.

(3) The element positions of commodity resource positions, card coupon resource positions and store resource positions identified by the S5 resource position module classification model can be directly used as commodity interactive controls, card and coupon interactive controls, and store interactive controls. output.

For the identified interactive control positions of all the above categories in the screen page, the system can click on any position of all the interactive operation point positions corresponding to the category or traverse all positions through the corresponding category configured by the user; The user wishes to execute the controls of a certain floor in the elevator navigation module, the coupon-type controls, etc. If there is no interaction operation point of the specified category in the screenshot of the screen page, a sliding operation of a certain step is performed on the page by default to reach the next screen position of the page, and then step S3 is repeated.

After the page is changed, the user can directly trigger the system to perform the automatic inspection of the current page after the change. The interactive operation mode of the page supports user configuration or default sliding, and the output inspection result is notified to the user in real time.

As an application example of the present invention, FIG. 2 is a schematic diagram of an application example of the present invention, and the input is a screenshot of a typical mobile terminal e-commerce promotion site page. After pre-classification and screening according to the aspect ratio, area, relative position and other characteristics of the DOM elements within the screenshot range, two types of ROIs of strip elements and resource bit elements are obtained respectively, and then the trained strip module is used to classify the model and resource bit. The module classification model identifies the ROI, and then obtains the ROI of the commodity resource position, the ROI of the card and coupon resource position, the ROI of the elevator, the ROI of the floor title, the ROI of the bottom tab navigation bar, and other types of ROI. After obtaining the floor title ROI, it can be further judged whether there is an empty floor problem; after obtaining the commodity resource position ROI, it can be further judged whether there is the same commodity type problem. At the same time, after obtaining the ROI of the elevator and the bottom navigation bar, this type of interactive operation point can be detected in the ROI area through OCR and template matching technology.

For those skilled in the art, the technical solutions described in the foregoing examples can be modified, or some technical features thereof can be equivalently replaced. All modifications and equivalent replacements made within the spirit and principle of the invention shall be included within the protection scope of the invention.

Claims (10)

1. An image-based on-end intelligent page quality inspection method is characterized by comprising the following steps:

s1: acquiring DOM element position information and page screenshots of a page, wherein the page screenshots comprise normal page screenshots, abnormal page screenshots and the like. The abnormal state screenshots comprise a page screenshot containing a hollow and a page screenshot containing a file overlapping coverage problem; and additionally acquiring DOM element position information corresponding to the empty pit area and the pattern overlapping coverage area.

S2: and respectively training a bar module classification model, a resource position module classification model, a pit anomaly detection model and a pattern overlapping and covering identification model according to the DOM element position information and the page screenshot obtained in the step S1.

S3: acquiring DOM element position information of a page to be inspected and a screenshot of the page to be inspected, and screening and intercepting strip elements to be inspected, resource location elements to be inspected and the like.

S4: and (4) classifying the to-be-inspected strip element and the to-be-inspected resource bit element obtained in the step (S3) respectively by using the strip module classification model and the resource bit module classification model obtained in the step (S2).

S5: and identifying different page problems, including empty pit type, empty floor type, empty screen type, repeated material type and pattern overlapping type.

2. The image-based intelligent end-on page quality inspection method according to claim 1, wherein the DOM element position information includes relative coordinates of a rectangular bounding box of the DOM element in the page screenshot, an aspect ratio of the rectangular bounding box, a ratio of each side corresponding to the rectangular bounding box to the screenshot, and an area of the rectangular bounding box.

3. The image-based intelligent end-to-end page quality inspection method according to claim 1, wherein in the step S2, the training process of the bar module classification model and the resource location module classification model specifically comprises: screening and intercepting bar-shaped elements and resource bit elements according to the DOM element position information and the page screenshot obtained in the step S1; training by adopting a deep learning residual error network or an automatic machine learning technology according to the strip element ROI and the corresponding category thereof to obtain a strip module classification model, inputting the strip element, and outputting the strip element as a strip element category; training by adopting a deep learning residual error network or an automatic machine learning technology according to the resource bit elements and the corresponding categories thereof to obtain a resource bit module classification model, inputting the resource bit elements, and outputting the resource bit elements as the categories of the resource bit elements;

the training process of the empty pit abnormity detection model specifically comprises the following steps: after S1 is executed in batch, firstly, a round of manual rechecking marking is carried out on the page screenshot containing empty pits obtained in S1 and the label of the bounding box coordinate of the DOM corresponding to the empty pit element, namely, a sample which is actually a non-empty pit label is deleted, an unmarked empty pit area is added, after the rechecked page screenshot containing the empty pits and the sample data of the marked bounding box thereof are obtained, an R-FCN target detection frame is adopted for training to obtain an empty pit type detection model, the input of the model is an image, and the output is whether the empty pits are contained or not;

the training process of the pattern overlapping coverage recognition model specifically comprises the following steps: and (4) after the page screenshot containing the case overlap coverage problem and the DOM element position information of the problem area are obtained according to the step S1, training by adopting an R-FCN target detection framework to obtain a case overlap coverage type detection model, inputting the page screenshot, and outputting whether case overlap coverage exists or not.

4. The image-based intelligent end-to-end page quality inspection method according to claim 3, wherein the screening condition of the strip-shaped elements is W/h e (5,6), W/W e [1, 3); the screening conditions of the resource site elements are h/W <3, W/h <3.5 and 15W < h W < 150W; w is the width of the element, h is the height of the element, and W is the width of the image in which the element is located.

5. The image-based intelligent end-to-end page quality inspection method according to claim 3, wherein the categories of the bar-shaped elements include elevator navigation modules, floor titles, bottom Tab navigation bars, and other categories; the categories of the resource allocation elements include a commodity resource allocation category, a card ticket category, a store category, and other categories.

6. The image-based intelligent end-to-end page quality inspection method according to claim 1, wherein the step S5 specifically comprises:

(1) for the empty pit type page problem, detecting whether empty pits exist in the page to be detected obtained in the step S3 by using the empty pit type abnormity detection model obtained in the step S2;

(2) for the problem of the empty floor page, judging whether the number of the floor title strip-shaped elements classified in the step S4 is more than or equal to 2 and the longitudinal relative distance between two adjacent floor title strip-shaped elements is less than a set empty floor threshold; if so, the page to be checked has an empty floor, otherwise, the page to be checked has no empty floor;

(3) for the problem of the blank screen page, taking the middle area of the screenshot of the page to be checked obtained in the step S3, extracting SURF characteristic key points in the middle area, if the number of the extracted SURF characteristic key points is smaller than a set key point threshold value, judging that the blank screen problem exists in the page to be checked, and otherwise, judging that the blank screen problem does not exist;

(4) for the problem of repeated materials, comparing the similarity of every two commodity resource level elements classified in the step S5, and judging whether an abnormal problem that different commodity resource level elements are associated with the same target object information exists or not, wherein the abnormal problem is that the similarity is greater than a set threshold value;

(5) and aiming at the problem of the page with the document overlapping coverage, detecting whether the screenshot of the page to be detected, which is obtained in the step S3, has the document overlapping coverage by using the document overlapping coverage identification model obtained in the step S2.

7. The image-based intelligent end-on page quality inspection method according to claim 6, wherein the height and width of the middle area in the step S5 are both greater than half of the height and width corresponding to the screenshot of the page to be inspected.

8. The image-based intelligent end-to-end page quality inspection method according to claim 6, wherein the similarity of every two of the commodity resource position class resource position elements classified in the comparing step S4 in the step S5 is specifically as follows: and (5) extracting the HOG features of the commodity resource position elements classified in the step (S4), and calculating the cosine similarity of the HOG features of every two commodity resource position elements.

9. The image-based intelligent end-on page quality inspection method according to claim 1, wherein the strip-shaped module classification model and the resource bit module classification model are used for identifying strip-shaped elements and resource bit elements of different categories in the page, and corresponding interactive control positions are set; configuring the position and the sequence of the interactive control required to be checked, simulating the position of the interactive control appointed in the clicked page according to the configuration, and acquiring the DOM element position information and the page screenshot of the clicked page; and if the appointed clicked position of the interactive control does not exist in the screenshot of the current screen, executing a downslide operation to reach the next screen position, continuing to simulate clicking the appointed position of the interactive control, and acquiring DOM element position information and the screenshot of the clicked page.

10. The image-based intelligent end-on page quality inspection method according to claim 9, wherein the strip-like module classification model and the resource bit module classification model are used to identify strip-like elements and resource bit elements of different categories in the page, and corresponding interactive control positions are set, specifically:

(a) the positions of the resource position elements of the identified commodity resource position class, the card ticket class and the shop class are directly used as the positions of the interactive controls of the commodity resource position class, the card ticket class and the shop class;

(b) recognizing the recognized strip-shaped elements of the elevator navigation module by adopting an OCR technology to obtain the position of an interactive control used for reaching a certain floor in the elevator navigation module, and recognizing by utilizing a template matching method to obtain the position of the interactive control of an arrow button used for expanding all contents in the elevator navigation module;

(c) and recognizing the strip-shaped elements of the identified bottom Tab navigation bar by adopting an OCR technology to obtain interactive control positions for jumping to other column pages.

Images