1 Introduction In modern automated production processes, machine vision systems are widely used in fields such as working condition monitoring, finished product inspection and quality control. Machine vision inspection systems typically use CCD (Charge Coupled Device) cameras to capture inspection images, convert them into digital signals, and then use advanced computer hardware and software technology to process the digital signals of the images to obtain various target image feature values. Based on this, they can realize multiple functions such as pattern recognition, coordinate calculation, and grayscale distribution maps. [1] In the quality inspection of printed materials, density detection and color detection based on test strips have been proven to be an effective quality control method. However, this quality control method mainly relies on the long-term experience accumulation and subjective judgment of the operator, and the repeatability and stability of the detection results cannot be guaranteed. [2] With the improvement of the automation level of printing machinery, the speed and sensitivity of printing quality inspection and control is also an inevitable trend of quality inspection and control. This requires that the detection of printing quality during the printing process can adapt to high-speed printing, accurately detect minor faults, and quickly feed back the detection information to the printing press. This paper applies the principles of machine vision to establish a full-screen printing quality inspection system. A CCD camera continuously photographs the printed materials, transmitting each frame to a local computer. Image processing software analyzes the image information, identifies images with quality problems, and identifies the corresponding quality issue. This information is then transmitted back to the operator or directly to the printing press for adjustments. This not only reduces the labor intensity of workers but also decreases defective products and improves production efficiency. 2. Full-Screen Printing Quality Inspection System Design 2.1 System Theoretical Design and Model The full-screen inspection system established in this paper mainly consists of four parts: image acquisition, image processing, data conversion, and result output. Image acquisition is composed of a CCD, lens, light source, video image acquisition card, and computer. Image processing mainly uses software programming to remove noise, perform geometric transformations, and locate the image. Data conversion converts the RGB data of the acquired digital image into printing feature values, i.e., the dot area ratio of CMYK ink. Result output mainly outputs the data calculated in the data conversion section and displays printing feature quantities, such as ink volume. The system mainly refers to the image acquisition equipment. The structural block diagram can be represented by Figure 1. The CCD, lens, light source and image acquisition card work together to complete the acquisition and digitization of the image. High-quality image information is the original basis for the correct judgment and decision of the system, and is the key to the success of the whole system. CCD devices can be divided into two categories: linear array and area array. [3] Linear array CCD can only obtain one line of image information at a time. The object being photographed must move in a straight line from the front of the camera to obtain a complete image. Therefore, it is very suitable for image detection of objects moving at a certain speed. Area array CCD can obtain the information of the entire image at a time. In the full-screen detection system, the Bayer-converted area array CCD of Sony is used. After experimentation, for small-format printed materials (200mm×200mm), when the object distance between the printed material and the lens is 15mm, the illumination of the light source is the most uniform and the imaging quality is good. Therefore, the lens focal length is 3.5-8mm, the imaging size is 1/3 inch, and the aperture is F1.4. The light source adopts the forward vertical illumination method. Image processing is the core of the full-screen detection system. It is equivalent to the human brain. With the rapid development of computer technology, microelectronics technology and large-scale integrated circuit technology, in order to improve the real-time performance of the system, many tasks of image processing can be completed with the help of hardware, such as DSP (Digital Signal Processing), dedicated image acquisition card, etc. [4] Software mainly completes the very complex, immature and still need to be explored and changed parts of the algorithm. Image digitization is implemented by the image acquisition card OK_MC10A of Jiaheng Zhongzi Company. It is a PCI bus-based acquisition card. When the optical image of the subject is imaged on the photosensitive surface of the CCD, the image data is stored in the computer memory in real time and displayed on the CRT (cathode ray tube) monitor. Single frame acquisition is used to adjust the optical image to the best state for post-processing. The image acquisition data is red, green and blue (RGB). Because the RGB data obtained by different devices are different, while printed matter is generally represented by four-color ink (CMYK) of cyan, magenta, yellow and black (CMYK). Directly using the RGB value of each pixel for comparison is a method of digital image analysis, but the accuracy of this method is not high. In the printing industry, ink volume control is a crucial aspect of print quality control, as it is closely related to printing characteristics such as solid density, dot gain, and printing contrast. Print quality is controlled by adjusting the ink volume of the four primary inks (magenta, cyan, and black) on the printed material, which in turn controls the ink volume of the printing press. Therefore, RGB data alone cannot directly determine print quality; converting the data into commonly used printing characteristics is necessary. This system primarily converts the data to the dot area ratio of the four primary inks. Offset printing presses typically have several ink zones. To facilitate future press control, the ink volume of the four primary inks is displayed in separate zones, with the ink volume of each zone calculated using the statistical average of the dot area ratio. 2.2 Software Design Upon entering the full-screen print quality inspection system software startup interface, the image acquisition device is turned on, and the printed image is displayed in real time. Single-frame capture is performed, followed by image processing, and the processed image result is displayed on the screen. Single-frame images are stored, RGB data is converted to CMYK data, and ink volume is displayed in separate zones. The software can be divided into the following modules: image acquisition and screen display, common image processing algorithms, image data conversion, and ink volume display. The image acquisition module primarily controls the real-time display of acquired images on the screen. Acquisition methods include single-frame acquisition and real-time acquisition and display. In the full-screen printing quality inspection system, we focus on preprocessing the acquired images before analyzing the image data, which is a relatively complex process. Therefore, we use a method where single-frame acquired images are stored in a designated location, and then the image is retrieved for processing using an image pointer. Common image processing algorithms include median filtering, geometric transformation, and localization. The acquired images are processed and the processing results are displayed via a menu-driven approach. Experiments show that using median filtering with a 3×3 template yields clear images and effectively removes noise, preparing for the next step of data conversion. Image data conversion completes the RGB to CMYK conversion. The acquired images are stored in bitmap format, with data format RGB24. This paper mainly uses a three-layer BP neural network (Back-Propagation Network) image recognition method to convert RGB values ​​to CMYK values. In the BP neural network, the color value of the image pixel is represented by RGB, so there are 3 input nodes, which input RGB values ​​respectively. The output is the color information value of the point, namely C, M, Y, K, so there are 4 output nodes, which output C, M, Y, K values ​​respectively. The selection of the number of hidden layer neurons is the key to the BP network. The experiment of this algorithm shows that the effect is better when there are 8 hidden layer units. [5] The convergence condition of the BP neural network requires that the initial input variables must take values ​​in the range of [0, 1], while the range of R, G, B as input values ​​in this system is 0 to 255. Therefore, all R, G, B values ​​need to be normalized. The task of ink volume display is to use the ink volume value of CMYK as the feature quantity of printing and display the ink volume in different areas on the screen. The full-screen detection system realizes the display of ink volume in different areas, which can correspond to the ink area on the printing press and realize the control of ink volume in the next step. 3 Experiment The main purpose of the experiment is to train a three-layer BP neural network and realize ink volume display. The training samples should be selected to represent the general rules of the represented object, which is beneficial for the neural network to summarize and generalize the rules contained in the samples. The minimum number of training samples should ensure that each output neuron appears at least once in the output value of these samples. In the experiment, 24 color patches of IT8.7/3 (CMYK) were selected as training samples and numbered in a certain order. Because four-color printing is generally used in actual printing, black ink must be considered, as the amount of black ink has a significant impact on the tonal reproduction and maximum contrast of the printed matter. These 24 color patches include various color rendering methods in four-color printing, including single-color, two-color, three-color, and four-color printing, with four-color printing being the primary method. This selection of training samples, compared with selecting only four-color overprinting samples, is also beneficial to improving the accuracy of the BP neural network. An EPSON 7600 simulated printing press was used to print color patches on offset paper with a basis weight of 90g/m2. The size of each color patch was 70mm×70mm. Through training, the weights and thresholds of the BP neural network were determined. The feasibility and accuracy of the algorithm can be verified by comparing C1M1Y1K1 calculated by the BP neural network with the standard CMYK. As shown in Figure 2, it can be seen that the errors of yellow ink and black ink are relatively large. The reason is that the CMYK standard data is electronic data, which does not take into account the dot gain, ink and paper characteristics, etc., and has a certain error with the actual CMYK value of the color block. In the vertical direction of the printing direction, the printed matter is divided into 8 ink areas to display the ink volume of each CMYK color. The ink volume is displayed using the dot area ratio obtained from the previous data conversion. The statistical average of the dot area ratio of each area is calculated as the ink volume of this area. The maximum value of the ink volume is 100%, that is, 1, which means that the ink is solid in this area. The programming is implemented using ActiveX control. [6] In the experiment, the ink volume of a single color block is displayed first, that is, the single training sample during data conversion is used for display. Then they are combined to display the ink volume. The ink volume of the standard image can be saved to establish a standard template. Figure 3 shows the ink volume display of four combined color blocks. 4. Conclusion The experiment shows that the accuracy of the BP neural network algorithm is improved because the training samples of the three-layer BP neural network consist of 24 color blocks in IT8.7-3CMYK, with a large proportion of four-color overprinted color blocks, and the addition of single-color, two-color, and three-color samples. The distribution of the halftone dot area ratio of the CMYK constituting the color blocks is mainly concentrated between 20% and 70%, which is commonly used in printing and can reflect the printing quality of the product. The ink volume display shows the ink volume distribution in different areas. In addition, offset paper was used in the experiment. If coated paper is used instead, the light source must be adjusted due to the different paper characteristics. The appropriate lighting method can be selected through experimentation. The full-screen quality inspection system can realize offline and online inspection of print quality. This paper mainly realizes the offline inspection of printed materials and displays the ink volume of printing feature values. Installing image acquisition equipment at different positions on the printing production line can realize online inspection of print quality. The authors' innovations include: using 24 color patches from ITIT8.7-3CMYK as training samples for a three-layer BP neural network, which improved the accuracy of data conversion; and realizing the display of ink volume in the samples. References [1] Liang Ji, Jiang Shiqin, Shen Liwei. Visual inspection system and its application [J]. Microcomputer Information, 2003, (12). [2] Gong Xiurui, Liu Xin. Real-time detection technology for printing quality [J]. Packaging Engineering, 2003, 24 (6): 45-49. [3] Wang Qingyou. CCD application technology [M]. Tianjin: Tianjin University Press, 2002. [4] William K. Pratt, translated by Deng Luhua, Zhang Yanheng, et al. Digital Image Processing [M]. China Machine Press, 2005, 179-183 [William K. Pratt. Digital Image Processing [M]. China Machine Press. 2005. 179-183 (in Chinese)]. [5] HORNIK K, STINCHCOMBE M, et al. Multilayer feed-forward networks are universal approximators [J]. Neural Networks, 1989, (2): 359-366. [6] Fu Xingwang, Liu Wangkai, Shen Weiqun. Design of a general-purpose curve control [J]. Ordnance Automation, 2006, 25(2): 91-92. Editor: He Shiping