1 Introduction Direct wafer bonding technology is a technology that involves directly bonding two mirror-polished wafers after surface cleaning and activation treatment at room temperature, followed by annealing to increase the bonding strength and form a whole. This technology does not require any adhesive, and the resistivity and conductivity of the two bonded wafers can be freely selected. The process is simple and it is the best means to prepare composite materials and realize micromachining [1]. It is often used in combination with other methods, which can both support and protect microstructures and realize electrical connections between mechanical structures or between mechanical structures and circuits [2]. When bonded wafers are used, they must first have good mechanical properties (void size and distribution and bonding strength), which is the basis for good electrical properties of bonded wafers [3]. The absence or minimal voids at the bonding interface are the original requirements for making reliable devices. There are two main categories of bonding detection methods: destructive and non-destructive. Currently, the most commonly used methods to describe the mechanical properties of bonding are image method, cross-sectional analysis method and bonding strength test. Image method is a non-destructive method and can be used for online real-time monitoring; while the latter two methods are destructive methods and require control modules. For silicon wafer bonding, infrared transmission, ultrasonic testing, and X-ray imaging are the three main imaging methods [4]. Although the spatial resolution of infrared transmission for detecting interface voids is not as good as that of ultrasonic testing and X-ray imaging, infrared methods have the advantages of being simple, fast, inexpensive, and readily available, and can be used directly in cleanrooms to obtain images of the bonded wafers before and after annealing. The other two imaging methods, although having high resolution, are expensive, time-consuming, and incompatible with cleanrooms, and cannot monitor the bonding process in real time. This paper mainly discusses the use of image processing technology based on the infrared transmission principle of wafers to overcome the disadvantages of high cost and technical complexity in previous testing methods. Taking silicon-silicon direct bonding as an example, an infrared detection device and related software modules were designed and built, and combined with a silicon wafer bonding device to achieve rapid and effective online monitoring of bonding processes and preliminary evaluation of wafer bonding quality. 2 Infrared Detection Principle The near-infrared part of light waves (wavelength about 0.75~1.5 μm) can pass through the wafer, and different wafers have different transmittance for infrared light. The minimum wavelength of infrared light that a wafer can transmit is shown in Table 1. If there is an unbonded area at the bonding interface of two wafers, the light will be reflected twice to form coherent light. When the CCD is taken, interference fringes will appear on the image. If the unbonded area is large and the gap height is small, many large interference fringes will appear. If the unbonded area is small, small Newton's rings will appear on the infrared image. When the gap at the bonding interface is large, infrared light can hardly pass through, and only black patterns will appear at the corresponding positions on the image. Therefore, the defect state and distribution of the bonded wafer can be successfully detected based on the infrared transmission image of the bonded wafer. However, if the monochromaticity of the light is not good, or the surface of the unbonded area is not very regular, Newton's rings cannot be observed. At this time, only the pattern of light and dark contrast can be observed on the image [5]. 3 System Design 3.1 Selection of Light Source and CCD The quality of the obtained image directly affects the complexity of the image processing program and the detection result. If the monochromaticity of the light source is better and closer to parallel light, the interference fringes of the image will be clearer and the quality will be better. However, monochromatic lasers or parallel light sources are bulky, while a key feature of infrared testing systems is their simple and compact structure. Providing narrow-band illumination is expensive and difficult to control; therefore, ordinary incandescent lamps are chosen as the light source. To obtain better infrared images, a double-sided mirror-polished silicon wafer is placed above the lens to filter out the influence of visible light on the image. Simultaneously, an ultra-low-light black-and-white camera, the WAT-902H, is selected, and its spectral response sensitivity curve is shown in Figure 1. Since the infrared wavelength is 750 nm to 1000 μm, a narrow-band infrared image can be obtained using an ordinary light source. Furthermore, because the camera's response sensitivity to the infrared band is not high, this can be overcome by increasing the light intensity, thereby obtaining a clear infrared interference image, preparing for subsequent image analysis and processing. 3.2 System Composition The structure of the testing system is shown in Figure 2, consisting of a light source adjustment device, a light source, a variable shading aperture, a test platform, a magnifying lens, a black-and-white CCD camera, a data acquisition card, and a computer. The light source and CCD are installed on opposite sides of the test sample, facing each other. The height of the light source is adjustable to adapt to the requirements of testing different wafers, thereby obtaining the clearest infrared images. A variable aperture is placed below the bonded wafer, with its central aperture adjustable from Φ1.8 to 50 mm. It controls the size of the light spot illuminating the bonded wafer. Generally, the inner aperture of the variable aperture is adjusted to match the size of the bonded wafer, but it can also be adjusted to be smaller to detect local features of the bonded wafer. The variable aperture optimizes the light source while simplifying the background of the infrared image, making the image outside the bonded wafer a single black, reducing the complexity of image processing and simplifying the system software. The light source illuminates the bonded wafer through the variable aperture. The light passes through the bonded wafer, through the lens, and is imaged on the camera, thus obtaining the infrared image of the bonded wafer. This image is sent to the computer via a data acquisition card, processed by the image processing program, and the test results are displayed. 3.3 System Software Module The hardware testing part of this instrument is connected to a PC. The acquired images are directly stored on the PC. Software can be used to process the images to obtain the required information, and it also provides image display and test result display functions. Using general-purpose office software such as Photoshop to process images requires the involvement of professional technicians specializing in bonding. Excessive human intervention directly impacts test results and is inconvenient. Therefore, we developed a corresponding software module using Visual C++, which allows for convenient and quick image processing and retrieval of necessary information without the need for professional personnel. The main processing module flow is shown in Figure 3. 3.3.1 Illumination Compensation Module As mentioned above, choosing a common incandescent lamp as the light source is the most economical and suitable option, but it also results in uneven illumination on the silicon wafer surface. Even within the bonding region, the center is brighter than the surrounding areas, and the gray values ​​of the bonding and unbonded areas at different illumination positions are very similar, which greatly complicates image segmentation. Therefore, an illumination compensation module was added, successfully solving the problem of uneven illumination. The illumination compensation curve was obtained through calibration fitting. 3.3.2 Contrast Enhancement Module: Due to the low contrast between bonded and unbonded regions in the image, image segmentation is difficult. Analysis of the grayscale histogram reveals that grayscale values ​​are concentrated in a certain range of 0-255. A contrast enhancement algorithm is employed to uniformly increase the difference between different parts, thereby widening the gap between bonded and unbonded regions and facilitating subsequent processing. This contrast enhancement algorithm differs from histogram equalization; it does not accumulate in the algorithm itself, and its effect is reflected in the histogram, evenly widening the spacing between various grayscale values ​​without changing the number of grayscale levels or their corresponding probability values. 3.3.3 Image Smoothing Module: During image acquisition, noise is inevitably introduced, making image smoothing an indispensable part of image preprocessing. The image smoothing algorithm used here is gradient-based, neutralizing the combined effects of mean filtering and median filtering. It suppresses noise while blurring interference fringes, thus laying the foundation for subsequent threshold segmentation. The implementation of the algorithm: Using a 3×3 neighborhood T[3][3], take the gray value gradient of the center point and its 8 neighboring points, divide T[3][3] into three regions according to the threshold T0, and the gray value of each point in the neighborhood is the gray value of its region. Use T[3][3] to traverse the whole image, and take the cumulative average of the gray values ​​of each point in the image. 3.3.4 Threshold segmentation module Threshold segmentation is the process of extracting the target object from the whole image. Here, it is to extract the area on the bonded part to prepare for the calculation of the bonding rate. 3.3.5 Bonding rate calculation module On the binary image after threshold segmentation, calculate the area of ​​the bonding region, and thus calculate the bonding rate. The bonding rate is defined as the percentage of the area on the wafer bonded to the total area of ​​the pre-bonded wafer. 4 Application of the system 4.1 Online monitoring of the bonding process Applying infrared detection to the bonding device can monitor the bonding process in real time. Figures 4(a) to (d) are four images taken during the pre-bonding process, clearly showing the propagation of the bonding wave from the center outwards. The numbers in the figures represent the pre-bonding time (the time from the initial contact between the two wafers after activation to the gradual bonding). After 1 minute, the bonding area basically no longer changes, and the pre-bonding result is shown in Figure 4(d). 4.2 Testing the wafer bonding quality The pre-bonded silicon wafers from the previous section were annealed at 120°C for 5 hours, and the infrared image of the bonded wafer is shown in Figure 5(a). The bonding quality is shown in Figure 5, with a bonding rate of 60.20%. Both sets of samples used single-sided polished p-type standard wafers. After cleaning and activation, they were bonded together and then annealed at low temperature to form a stable bond, obtaining test samples. The bonded wafers were placed on the testing instrument for testing, and corresponding image processing was performed to obtain the infrared images of samples 1 and 2, as shown in Figures 6 and 7. In Figures 6(a) and 7(a), the circular areas are the wafers to be bonded, with the bright areas representing bonded areas and the dark areas representing unbonded areas. The area outside the circular areas (i.e., the bonded wafers) is the image background. The distribution, number, and size of the voids (unbonded areas) can be roughly observed from the figures. To obtain more accurate data, image (a) was processed using image processing software, including contrast enhancement, smoothing, and segmentation, resulting in Figures (b) and (c), thus yielding the bonding rate. The bonding rate of Sample 1 was 28.12%, which is very low, and the distribution of voids in Figure 6 shows that Sample 1's bonding was poor. The bonding rate of Sample 2 was 66.12%. Comparing Figures 6 and 7, the bonding quality of Sample 2 is significantly better than that of Sample 1, indicating that thinner bonding wafers are easier to bond. This bonding phenomenon can be understood from the plate theory: the thinner the wafer, the less force is required on the interface surface to overcome the warping of the wafers and bond them together. On the right side of Figure 7(a), a circular dark area with clear circular interference fringes can be seen. This is because there was a particle contamination on the wafer at this location before bonding. Therefore, it can be inferred that voids formed by particle contamination at the interface appear as relatively regular circular dark areas in the infrared image. 5. Conclusion The infrared detection system for wafer bonding quality developed in this paper has advantages such as low cost and simple implementation principle and method. Using this detector, the bonding rate and defect distribution of the wafer can be quickly obtained, thereby achieving rapid evaluation of wafer bonding quality. Analysis and comparison of the bonding quality of bonded wafers under different process conditions, including bonding rate and void distribution, combined with parameters such as bonding strength, can help understand the mechanism of wafer bonding, thereby guiding the bonding process and optimizing process parameters. This detector is more flexible and practical, and can be used not only for quality inspection of bonded wafers made of homogeneous materials, but also for inspection of bonded wafers made of dissimilar materials, for screening suitable bonded wafers for further process research, etc. Meanwhile, this detector, integrated into the bonding device, can monitor the dynamic images of the bonding process in real time, observe the bonding waves, and provide real-time guidance for the bonding process. However, the currently implemented software functionality is still very basic, only allowing for a preliminary assessment of bonding quality. To obtain more information about the bonded wafers, additional software functionality is needed; this is a shortcoming of the system and an area worthy of improvement.