Computer-implemented process to enhance edge defect detection and other defects in ophthalmic lense

Abstract

To provide a computer-implemented method to train an inspection system for defect inspection of contact lenses.SOLUTION: A method identifies defects around an edge of a contact lens by training a contact lens inspection system for classifying the defects with different criteria, configured to: capture a high resolution image; extract a circular edge of a lens; transform the circular edge to a horizontal line representing the circular edge; restrict an image size by eliminating pixel information around the edge; divide a horizontal edge image into overlapping portions; and stack the extracted images vertically to form a single high-resolution image. This is ideal to be processed and analyzed by convolution neural networks after augmenting an original image dataset with new images generated by generative adversarial networks, to enable accurate classification of the defects.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention extracts specific relevant pixel data and removes irrelevant data to accurately By increasing the efficiency of the image classification process in terms of accuracy and speed, It involves achieving and maintaining high-resolution images of the Helps in rapid training of inspection systems and during contact lens inspection in automated systems Increased efficiency is achieved through realignment and reconstruction of images applying trained decision-making models. Successful. [Background technology]

[0002] The contact lens defect inspection system and method is directed to increasing efficiency while reducing costs. It continues to evolve towards the future. Pattern recognition, image correlation, histogram equalization, dithering, etc. Below are some common image processing algorithms commonly used in inspection methods. As defect inspection standards become more stringent, additional work is being done to improve inspection efficiency at the expense of time. Additional algorithms are implemented. Every additional algorithm increases the inspection time and Reduces productivity. CPU speed, advanced GPU (Graphics Processing Unit) Technological improvements in image processing speeds, such as image processing units and high-speed memory, have helped improve image processing speeds. The images acquired from high-resolution cameras are getting larger and larger, reducing productivity. Therefore, high accuracy is required to support the inspection process in achieving high accuracy and repeatability. It is most important to seriously consider software methods for reconstructing resolution images. Summary of the Invention [Problem to be solved by the invention]

[0003] Neural networks improve the quality and productivity of your inspection system. It offers new means to improve inspection quality, but it also significantly increases image size. Therefore, to minimize pixel data that leads to degradation of important defect data, image compression is Therefore, the quality of the inspection is affected and the efficiency of the next process is affected. Neural networks are more efficient for medium-sized images. Increasing image size negatively impacts the performance of neural networks and deep learning methods. It has an effect.

[0004] Current technology is limited in its ability to identify and extract features from small defects, especially when processing high-resolution images. If it is a fundamental requirement of the system, the software method can be used without compromising inspection quality. Lack of use. [Means for solving the problem]

[0005] The present invention utilizes a high-resolution camera, supported by a high-performance CPU, and non-critical pixel data. Identify and discard the regions, and keep only the regions that are important for detecting defects (i.e., edges). A fast application complemented by a GPU to analyze and process images optimized by It advocates the use of computer-implemented processes by incorporating accessible memory. The computer-implemented process is not compressed, but is pre-processed to minimize redundant pixel data. The uncompressed image is processed and smoothing techniques are applied to highlight defect candidates, thus improving detection. When identifying defects in lenses, high resolution images are , a deep neural network consisting of several neural network modules for feature extraction and classification. The layer learning module generates a number of optimal and predetermined size images that are more suitable for high-speed processing. Even in the case of defect identification in a lens, the images are relocated or split into separate images from the original image. It is important to note that no image compression is applied when extracting.

[0006] One advantage of shrinking or rearranging the image is that it eliminates redundant pixels around the edges of the lens. Subsequently, deep learning modules are used for enhanced defect detection in terms of speed and accuracy. The image is simultaneously divided into predetermined image sizes that can be easily processed by the module.

[0007] One of the objectives of the present invention is to provide a contact lens display without compressing or distorting pixel data. The objective of the present invention is to provide a high-resolution and optimized pre-processed image of the circular edge of the image.

[0008] Another aspect of the present invention is to use the circular edge of the contact lens to aid in faster processing of the image. This is done to detect the edges of a circular contact lens. This is achieved by extracting the edges, unwrapping the edges, and converting them to horizontal edges. To minimize the image size, redundant pixel data around the edges is removed. This will be further optimized.

[0009] One advantage of arranging pixels as horizontal and vertical edges is that it improves processing, which is It helps the algorithm to process images faster, i.e. the processing speed is significantly improved. The algorithm may be an edge detection algorithm.

[0010] Vertically stack images (image area segments) of a given size to form a square image. Another advantage of overlapping is that the length and width of the square images are the same. For example, Scanning a square image requires fewer pixels per pixel than scanning a rectangular image. The advantage is the speed of the calculation, not the number of calculations. For example, it is significantly faster to process a square image compared to a rectangular image.

[0011] Another object of the present invention is to further refine horizontal edges that are divided into shorter lengths that overlap each other. and stack them on top of each other to suit the requirements of the input layer of any neural network. To achieve a square size, a square image is generated with all edge data matching the square size. The goal is to avoid cluttering the image with redundant black pixel data.

[0012] Another object of the present invention is to provide a computer-implemented process during the construction of a testing system for machine learning. The goal is to create a group of analytical modules to aid in the training of process.

[0013] Another object of the present invention is to generate some new realistic defect features that are similar to the original defects. To further transform the segmented images, we use a generative adversarial network algorithm. The newly generated images are then used in the construction of an inspection system for machine learning. , will be used to further enhance training in computer-implemented processes.

[0014] Other aspects of the invention include various combinations of one or more of the foregoing aspects of the invention, such as and the invention as seen in or as can be derived from the following detailed description. The foregoing aspects of the invention include any one or more combinations of the various embodiments described herein. It should be understood that there are corresponding computer-implemented processes that are also aspects of the invention. Other embodiments of the invention are described in the following detailed description of specific embodiments of the invention, and in the accompanying drawings, in which: It is understood that the invention may be derived by those skilled in the art from both the description of the system and the specific embodiments. It should be understood. [Brief explanation of the drawings]

[0015] Certain features, aspects, and advantages of the present invention are set forth in the following description, appended claims, and accompanying text. The present invention will be better understood with reference to the accompanying drawings, in which: [Figure 1] 1 is a flowchart illustrating steps for training a contact lens inspection system to identify defects around the edge of a contact lens and classify them by different criteria to implement one of the embodiments of the present invention. [Figure 1a] 10 is a flowchart illustrating steps for training a contact lens inspection system to identify defects in contact lenses and classify them by different criteria for practicing another embodiment of the present invention. [Figure 2] FIG. 10 is an explanatory diagram of an image of a contact lens edge. [Figure 2a] FIG. 10 is an explanatory diagram of an image of the contact lens edge after polar coordinate transformation. [Figure 3] 2b shows images of the edge extracted regions of FIG. 2a stacked in a particular order. [Figure 4] This is an illustration of a high-resolution image of a contact lens suitable for inspecting defects within the lens. [Figure 5] FIG. 5 is a diagram of the image of FIG. 4 after identifying the region to be extracted. [Figure 6] FIG. 6 is a view of the image of FIG. 5 after extraction of the individual overlapping regions to be used as input to the next process. DETAILED DESCRIPTION OF THE INVENTION

[0016] The following description of preferred embodiments of the present invention forms a part of the present specification and is intended to illustrate, by way of example, the principles of the present invention. Reference is now made to the accompanying drawings, which show, by way of illustration, specific embodiments which may be used. Other embodiments may be utilized and structural changes may be made without departing from the scope of the invention. I want you to understand that this is possible.

[0017] A general flow chart of the system and method according to the present invention is shown in Figure 1. System 1 0 is to acquire a high resolution image of the contact lens in the acquisition process operation 12. Next, in an extraction process operation 14, the acquired image is processed to extract the lens The extracted circular edges are then transposed in operation 16. In the removal process operation 18, the watermark is removed to minimize the size of the image. The irrelevant pixel data around the flat edges is removed. Next, in the segmentation process operation 20 The horizontal edge image is then divided into several overlapping images. are stacked one on top of the other to facilitate analysis in stacking process step 22. A high resolution image of the contact lens edge is generated to allow for accurate detection of the contact lens. The reconstructed and realigned images in step 22 are then used in deep learning for training and analysis. This can be used as input to the learning module. The process flow ends at step 24.

[0018] A general flow chart of another embodiment of the system and method according to the present invention is shown in FIG. 1a. 3. The system 30 acquires a high resolution image of the contact lens in an acquisition process operation 32. The image is then processed to the positioning process operation 33. The lens circular edge is positioned in the drawing process step 34. The circular edge of the contact lens is drawn to closely surround the circular edge of the lens. The pixel data outside the area is filled with dark pixels in a fill process step 35. The predetermined area is The contact lens image may be 16x16 to 128x128 pixels. In a programmable division process operation 36, the data is divided into predetermined sizes and overlapped. The predetermined sizes may be equal sizes. Alternatively, the predetermined sizes may be different sizes. In a deep learning extraction process operation 38, the marked regions in the image are are extracted, stored separately, and used as input to the deep learning module for training and analysis. The process flow ends at step 40.

[0019] Having described the general system and method according to the present invention, the following paragraphs This provides details of the process operation described above.

[0020] FIG. 2 is a pictorial representation of the process flow chart of FIG. 1. In FIG. 2, lens 5 0 represents a high resolution image of a contact lens. In Figure 2, defect 51 is a tear defect. This represents a concept of unwrapping the circular edge of a lens that will be described below and will be understood by those skilled in the art. The lenses 50 are strategically positioned so that they are aligned from A1-A2 as shown in FIG. This image is unwrapped or transposed into a full horizontal image. The outer circle in Figure 2 is the pixel area segmentation area in Figure 2a. As shown by points 500, 502, 504, 506, and 508, The pixel regions can be of equal size or of different (but equal) size. The image area segment 500 in FIG. 2a may have a position as shown in FIG. It starts before A1-A2 and ends after position 52, which overlaps image area segment 502. Image region segment 502 of 2a starts before position 52 of image region segment 500; It ends after location 53 where it overlaps image region segment 504. Point 504 begins before position 53 of image region segment 502 and 2a and ends after location 54 where it overlaps with image region segment 6. The image area segment 508 begins before position 54 in the image area segment 504 and overlaps position 5 5. Image region segment 508 in FIG. 2a ends at the position of image region segment 506. 55 and ends after the position A1-A2 that overlaps the image area segment 500. One advantage of this method of overlapping adjacent segments is that it minimizes the risk of image edges. This ensures that no surrounding areas are lost or omitted. It is important to note the location of defect 51 in the image. In FIG. 3, image 58 is The image region segments include region segments 500, 502, 504, 506, and 508. The points 500, 502, 504, 506, and 508 are vertically aligned due to overlap. Image area segments 500, 502, 504, 506, 50 8 are stacked one on top of the other as shown in FIG. 3 to form a square image 58. The number and width of segments in the ment are automatically adjusted to produce a square image when stacked. The square shape of image 58 is determined by subsequent process steps. It is important to note that subsequent process steps may result in different configurations of the image shape or The same is taken into consideration while placing the image if the size required. The image in Figure 3 is a reconstructed high-resolution image of the intact lens edge. Deep learning modules comprise neural networks that aid in rapid training and analysis. Again, the tear in the stacked image 58 shown in FIG. It is important to note the location of depression 51.

[0021] An image of the contact lens is shown in Figure 4. Figure 4 is a diagram of the process flow chart of Figure 1a. In FIG. 4, 66 is a contact lens encapsulated within a square image 65. The outer edge of the contact lens is first detected, and then the outer boundary Then, fill in the dark pixels between the outer boundary and the outer edge of the lens. The area enclosed by the circular edge of 66 and the square 65 is not important for inspection, so Filled dark pixels. Lens edges are identified by edge detection algorithms. and the high-resolution images of the contact lenses in Figure 4 with sizes X1 and Y1 in Figure 5 are The image is divided into multiple overlapping image segments. The images may be of equal size. Image segmentation can be achieved in several ways: One way is to define and draw the boundary of the image, and once the boundary is drawn, Once the size of the image is calculated, the size of the entire image can be calculated. The high-resolution images in Fig. 5 are 600, 602, 604, 606, 608, 610, and 61 Each divided image is preferably divided into nine equal parts: 2, 614, 616. Suitable for high-speed processing using graphics processing units and deep learning software modules The divided image sizes X' and Y' are used to calculate the neural network. Before training, a predetermined set of parameters is used to accommodate better analysis of defect features and increased efficiency. A typical segmented image of size X' and Y' is shown at 70 in Figure 6.

[0022] While several embodiments of the present invention have been described herein, the foregoing is merely illustrative. It will be apparent to those skilled in the art that these are provided by way of example only and not by way of limitation. Numerous modifications and other embodiments are within the scope of the art and are within the scope of the appended claims. and equivalents thereof.

Claims

1. 1. A computer-implemented method for training an inspection system for defect inspection of contact lenses, comprising: obtaining an uncompressed, high-resolution image of the contact lens; preprocessing the contact lens image by applying a smoothing algorithm to enhance defect information without compressing pixel data; Segmenting the pre-processed contact lens image into square regions formed from the image with overlap regions obtained from the pre-processed contact lens image; applying a generative adversarial network algorithm to transform the overlapped images to generate a number of defect features similar to the original defects in the contact lens images; and using the generated number of defect features to train the inspection system for contact lens defect inspection.

2. The computer-implemented method of claim 1 , further comprising removing redundant data that is not related to the defects being inspected.

3. Segmenting the preprocessed contact lens image includes: extracting a circular edge of the contact lens image, and horizontally transposing the extracted circular edge to join two adjacent segments, such that there is an overlapping area of ​​the adjacent segments of the extracted circular edge; Dividing the horizontally displaced extracted circular edge into a plurality of segments; and stacking the plurality of segments sequentially.

4. Segmenting the preprocessed contact lens image includes: The computer-implemented method of claim 1 , comprising dividing the preprocessed contact lens image into overlapping squares.

5. The computer-implemented method of claim 4 , further comprising filling redundant pixels outside the circular edge with black pixels.

6. The computer-implemented method of claim 1 , wherein the size of the segments is preset to be easily processed by a deep learning module for enhanced defect detection.

Images