Microscopic view field splicing method with large view field, high spatial resolution and low cost
Through the image stitching technology of sliding window and zigzag mode, combined with the feature fusion of neural network and Transformer module, the problems of error accumulation and low feature matching accuracy in microscopic image stitching are solved, and large field of view, high spatial resolution and low cost microfield of view stitching are achieved, improving the splicing accuracy and speed.
Patent Information
- Application Number
- CN202510268769.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-07
- Publication Date
- 2025-07-22
AI Technical Summary
The existing microscopic image stitching methods have problems with error accumulation and low feature matching accuracy, which makes it difficult to achieve large field of view, high spatial resolution and low cost microscopic field of view stitching.
The sliding window strategy and zigzag pattern are used for image stitching, combined with neural network feature extraction and global feature fusion of Transformer module, feature matching is used using mutually close neighbor algorithm, and the final stitching image is generated through affine transformation and Laplace pyramid smoothing technology.
It significantly improves the accuracy and splicing accuracy of feature matching, reduces error accumulation, and realizes large field of view, high spatial resolution and low cost micro field of view splicing, with a wide range of application and fast operation speed.
Smart Images
Figure CN120355566A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of optical imaging and microscopy technology, and particularly relates to a microscopic field-of-view stitching method with large field-of-view, high spatial resolution, and low cost. Background Art
[0002] Microscopy technology plays a crucial role in the biological and medical fields, providing the ability to explore the microscopic world for diverse applications. However, the inherently limited field of view (FoV) of microscope objectives poses challenges in capturing a comprehensive view of samples or rare events. To alleviate this limitation, all-field imaging technology has emerged as a promising approach. Among the strategies of all-field imaging, the method based on field-of-view scanning and stitching is the most widely adopted technique. Specifically, this technique involves scanning the sample to obtain a series of partially overlapping images, and then seamlessly stitching these images together to form the final large-field-of-view sample image. To further expand the field of view of microscopic imaging, techniques such as lensless holography, ptychography, and gigapixel imaging using multi-camera arrays have been further proposed and applied.
[0003] Lensless holography is a unique method that avoids the need for lenses or other optical elements and enables dynamic imaging over a large field of view. However, lensless holography has a low spatial resolution and a fixed field of view size that cannot be extended. In contrast, Fourier ptychography has significantly improved imaging quality compared to lensless holography. This technique not only expands the field of view but also significantly improves the spatial resolution, thus effectively increasing the Spatial Bandwidth Product (SBP). However, the field of view of Fourier ptychography is still limited by the micro-objectives used. To overcome this limitation, the ptychographic iterative engine was proposed and applied in practice. The ptychographic iterative engine directly expands the field of view by utilizing a series of diffraction patterns corresponding to different scanning regions. This method not only avoids aberration but also achieves diffraction-limited spatial resolution. However, it requires intensive scanning, which increases the complexity of the operation. Gigapixel imaging technology based on multi-camera arrays provides the ability to capture high-resolution images with an extremely large field of view. This method eliminates the need for mechanical or electronic scanning required by ptychography, but this technique requires a large number of cameras, resulting in a significant increase in cost. When applied to full-field imaging, techniques such as lensless holography, ptychography, and gigapixel imaging based on multi-camera arrays have poor timeliness and high time costs. Therefore, there is still a lack of an ideal micro-field-of-view stitching technique that can achieve a large field of view, high spatial resolution, and low cost. Summary of the Invention
[0004] Aiming at the disadvantages of error accumulation and low feature matching accuracy in existing micro-image stitching methods, the present invention provides a micro-field-of-view stitching method with a large field of view, high spatial resolution, and low cost, which can perform full-field micro-imaging of the sample to be measured and helps biologists and medical researchers observe and analyze samples over a larger range.
[0005] A micro-field-of-view stitching method with a large field of view, high spatial resolution, and low cost, the micro-field-of-view stitching method comprising the following steps:
[0006] S1, obtaining a series of small-field-of-view images of the sample to be measured covering a large field of view range and constructing a small-field-of-view image matrix;
[0007] S2. For the small field of view image matrix, first, the small field of view images in the first row and the first column are sequentially spliced to form a reference image. Then, in a zigzag pattern, the remaining small field of view images are sequentially spliced using a sliding window. When splicing, the sliding window is paired with the image to be spliced, and multi-dimensional feature matching is performed on the horizontal and vertical features within the sliding window and the image to be spliced. The sliding window is constructed based on multiple small field of view images containing image information.
[0008] S3. Based on a neural network, after the two images to be spliced paired with the sliding window are spliced along the third dimension, feature extraction is performed to generate a feature matrix, and the coarse-grained local features of the images to be spliced are output.
[0009] S4. Based on the Transformer module, the coarse-grained local features output by the neural network are converted into global features. Using two-dimensional position encoding technology, combined with self-attention layers and cross-attention layers, a confidence matrix is generated through activation by the Softmax function.
[0010] S5. Based on the confidence matrix and the extracted global features, the mutual nearest neighbor algorithm is used for feature matching. The random sample consensus method is adopted to accurately pair the identified features to form a set of feature pairs.
[0011] S6. Based on the set of feature pairs, the images to be spliced are aligned through affine transformation, and the relative displacement between the target microscopic image formed by splicing multiple images to be spliced and the sliding window is calculated. The image blending technology with Laplacian pyramid smoothing is used to iteratively downsample the images to be spliced to improve the convergence of chrominance in the overlapping area and generate the final spliced image.
[0012] Step S1 further includes:
[0013] S11. Place the sample to be measured on the translation stage and adjust the position of the microscope objective lens so that the sample presents a clear in-focus image on the image sensor.
[0014] S12. Control the position of the translation stage where the sample is located to change the sample position. Each time the sample position is changed, the image sensor is used to take a picture of the small field of view area corresponding to the sample, and the sample is scanned to obtain a series of small field of view images covering a large field of view range.
[0015] Further, in step S2, the process of constructing the sliding window is as follows: Select three adjacent small field of view images to be spliced and add a zero-value image for combination. Then, with the center of the combined image as the reference, a region with the same size as the image to be spliced is cropped as the sliding window image.
[0016] Further, in step S3, the neural network includes an input convolutional block, a first residual convolutional block, a second residual convolutional block, a third residual convolutional block, a first upsampling layer, and a second upsampling layer connected in sequence;
[0017] The input convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence; the kernel of the convolutional layer of the convolutional layer is 7×7, the stride is 2, and the padding is 3; the input convolutional block performs preliminary feature extraction on the spliced image;
[0018] The first residual convolutional block, the second residual convolutional block, and the third residual convolutional block each include two convolutional layers, two batch normalization layers, and two ReLU activation layers. Among them, the convolutional layers of the first residual convolutional block, the second residual convolutional block, and the third residual convolutional block all use 3×3 kernels, the padding is 1, and the strides are 1, 2, and 3 respectively; the first residual convolutional block, the second residual convolutional block, and the third residual convolutional block further extract the features extracted by the input convolutional block, and the extracted features are then successively passed through the first upsampling layer and the second upsampling layer for two-step upsampling to output the key features of the microscopic image. The upsampling operation is implemented using a transposed convolutional layer with a kernel size of 1.
[0019] Step S4 further includes:
[0020] The Transformer module uses two-dimensional position encoding technology to encode rough features, and then injects position awareness to improve the matching accuracy of blurred regions; the position-encoded features sequentially traverse the self-attention layer and the cross-attention layer to enrich feature extraction; among them, the self-attention layer maintains a consistent input, and the cross-attention layer adjusts the dynamic input according to the direction of attention; after the outputs of the self-attention layer and the cross-attention layer are activated using Softmax, the two resulting activation matrices are multiplied to obtain a confidence matrix.
[0021] Further, in step S6, the process of generating the final stitched image is as follows: First, the Laplacian pyramid technology is used to smooth and iteratively downsample the overlapping regions of the two images to be stitched to accelerate the convergence of the chromaticity consistency in the overlapping regions; then, Gaussian pyramid transforms are respectively applied to the edges of the two overlapping images to calculate their respective overlapping weights. Finally, these overlapping weights are mixed to generate a high-quality final stitched image.
[0022] Further, the small field of view images include fluorescence microscopic images and bright field microscopic images.
[0023] The beneficial effects of the present invention are:
[0024] The present invention has a microscopic field of view stitching method with a large field of view, high spatial resolution and low cost. Through a sliding window strategy, the features of multiple adjacent images are fully utilized to reduce the accumulation of single-direction errors, and small-field microscopic images are seamlessly merged into full-field images. The framework uses deep learning to extract features and match them, which significantly improves the accuracy of feature matching in the paired stitching process. The neural network used in the present invention is trained in an unsupervised manner and is not limited to the modality of microscopic imaging, thereby improving the practical applicability of the method. Image mixing is used to eliminate the effects of changes in contrast, color, brightness and position during scanning. Operator fusion technology can increase the reasoning speed of the neural network and improve real-time performance. Compared with other solutions, the present invention has the advantages of high stitching accuracy, wide application range, and fast running speed, thereby providing a new technical means for full-field microscopic imaging. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 The present invention is a flow chart of a microscopic field of view stitching method with a large field of view, high spatial resolution and low cost;
[0026] Figure 2 A schematic diagram of stitching a full-field fluorescence microscopic image obtained by applying the stitching method of the present invention, including: (a) a small-field fluorescence microscopic image obtained by scanning, (b) a stitched full-field fluorescence microscopic image, and (b1) an enlarged view of the stitched full-field image area;
[0027] Figure 3 A schematic diagram of stitching a full-field bright-field microscopic image obtained by applying the stitching method of the present invention, including: (a) a small-field bright-field microscopic image obtained by scanning, (b) a stitched full-field bright-field microscopic image, and (b1) an enlarged view of the stitched full-field image area. DETAILED DESCRIPTION
[0028] The following examples will enable those skilled in the art to more fully understand the present invention, but are not intended to limit the present invention in any way.
[0029] The present invention discloses a microscopic field of view stitching method with large field of view, high spatial resolution and low cost, and the microscopic field of view stitching method comprises the following steps:
[0030] S1, obtaining a series of small-field-of-view images of the sample to be tested covering a large field-of-view range, and constructing a small-field-of-view image matrix;
[0031] S2. For the small field-of-view image matrix, first, the small field-of-view images in the first row and the first column are spliced together in sequence as the splicing reference. Then, using the sliding window technique, the remaining small field-of-view images are spliced in a zigzag pattern in turn. During the splicing process, the sliding window is paired with each small field-of-view image to be spliced, and multi-dimensional matching is performed by comparing the horizontal and vertical features within the sliding window with the features of the image to be spliced. This sliding window is constructed based on multiple small field-of-view images containing image information;
[0032] S3. Based on the neural network, after the two images to be spliced paired with the sliding window are spliced along the third dimension, feature extraction is performed to generate a feature matrix, and the coarse-grained local features of the images to be spliced are output;
[0033] S4. Based on the Transformer module, the coarse-grained local features output by the neural network are converted into global features; using two-dimensional position encoding technology, combined with the self-attention layer and the cross-attention layer, a confidence matrix is generated through activation by the Softmax function;
[0034] S5. Based on the confidence matrix and the extracted global features, the mutual nearest neighbor algorithm is used for feature matching; the random sample consensus method is adopted to accurately pair the identified features to form a set of feature pairs;
[0035] S6. Based on the set of feature pairs, the images to be spliced are aligned through affine transformation, and the relative displacement between the target microscopic image spliced by multiple images to be spliced and the sliding window is calculated; the image mixing technology smoothed by the Laplacian pyramid is used to iteratively downsample the images to be spliced to improve the convergence of chrominance in the overlapping area and generate the final spliced image.
[0036] As Figure 1 shown, the specific steps of a microscopic field-of-view splicing technology that can achieve large field-of-view, high spatial resolution and low cost in the present invention are:
[0037] (1) Acquisition of small field-of-view images. Place the sample to be measured on the sample stage, move the microscopic objective lens up and down so that the sample presents a clear in-focus image on the image sensor, and use the image sensor to obtain a fluorescence microscope; precisely control the position of the translation stage where the sample is located. Each time the position of the sample is changed, the image sensor takes a picture of the small field-of-view area corresponding to the sample. After the sample scanning is completed, the small field-of-view images to be spliced are obtained in sequence. In this embodiment, Figure 1 Step 1 in shows the acquisition sequence of 25 images to be spliced. In other scenarios, the acquisition methods of 9, 16 or other numbers of images are the same.
[0038] (2) Sliding window stitching strategy. The window-based field of view stitching strategy aims to minimize error accumulation. The specific stitching process includes: sequentially stitching the images in the first row and the first column. Both the reference image and the image to be stitched are small field of view images collected. The remaining images are stitched in a zigzag pattern in turn. The reference image is the sliding window, and the image to be stitched is the small field of view image. The sliding window is constructed based on the small field of view image containing the information of multiple images. Combine three adjacent images of the image to be stitched and a zero-value image, and crop the combined image along the center to obtain a sliding window image with the same size as the image to be stitched. During the zigzag stitching process, the sliding window is paired with the image to be stitched, and the horizontal and vertical features within the window are matched with the image to be stitched in multiple dimensions to improve the calculation accuracy of the stitching position and reduce the error accumulation in a single direction.
[0039] For this embodiment, this strategy forms a new matrix with the already stitched images 1, 6, 2 and a zero-value image, and then crops the matrix along the center. The cropped part is the sliding window. The sliding window encompasses the information from all three images. By utilizing the horizontal and vertical features within the sliding window, additional constraints are introduced for the calculation of the stitching position, thereby facilitating a more accurate calculation of the stitching position and effectively preventing the accumulation of unidirectional errors.
[0040] (3) Feature extraction. Using a neural network can improve the quality of feature extraction and alleviate the problems of pairing failure or pairing error.
[0041] This patent introduces a specially designed neural network structure to extract high-quality features. The backbone of the neural network consists of four convolutional blocks and two upsampling layers. Using a pair of images with a size of 1024×1024×3 as input, before inputting the merged data into the initial convolutional block, they are first stitched along the third dimension. The initial convolutional block contains a convolutional layer with a kernel of 7×7, a stride of 2, and a padding of 3, as well as a batch normalization layer and a ReLU activation layer. Subsequently, the extracted features pass through three residual convolutional blocks, each block having two convolutional layers (using 3×3 kernels, padding of 1, and strides of 1, 2, and 3 respectively), two batch normalization layers, and two ReLU activation layers. Finally, a fine-grained feature matrix is obtained through two-step upsampling, and the upsampling operation is implemented using a transposed convolutional layer with a kernel size of 1. This network is trained on the large-scale ImageNet dataset, has strong generalization ability, is applicable to multi-modal microscopic images, can efficiently and accurately extract key features from microscopic images, and effectively alleviates the problems of pairing failure or error.
[0042] (4) Confidence matrix calculation. The present invention utilizes the global feature fusion technology based on Transformer to convert the coarse-grained local features into global features, exceeding the receptive field limitation of convolutional neural networks. The Transformer module uses two-dimensional position encoding technology to encode the rough features, thereby injecting position perception and significantly improving the matching accuracy in the fuzzy area; the position-encoded features sequentially traverse the self-attention layer and the cross-attention layer to enrich feature extraction. The self-attention layer maintains a consistent input, while the cross-attention layer adjusts the dynamic input according to the direction of attention; the outputs of the self-attention layer and the cross-attention layer use the Softmax activation, and the two activation matrices are multiplied to obtain the confidence matrix. In this step, by fusing the self-attention layer and the cross-attention layer, the global neural network can generate dense pixel matches. By combining two-dimensional position encoding with traditional position encoding technology to encode the rough features, position awareness is injected, significantly improving the matching accuracy in the fuzzy area. These position-encoded features will sequentially pass through the self-attention layer and the cross-attention layer, thus enriching the extracted features.
[0043] (5) Feature matching and screening. The confidence matrix is processed by SoftMax to obtain the soft mutual nearest neighbor matching probability for initial matching. The feature matching results are calculated based on the confidence matrix and the mutual nearest neighbor algorithm, realizing the accurate calculation of feature matching. This step uses random sample consensus to accurately pair each group of features identified at their respective best positions, thereby obtaining the corresponding feature pairs and improving the accuracy and robustness of stitching.
[0044] (6) Pairwise image stitching. The images are aligned through affine transformation, and the relative displacement between the target microscopic image and the sliding window is calculated to ensure seamless stitching.
[0045] (7) Image blending. The image blending technology using Laplacian pyramid smoothing iteratively downsamples the overlapping images to improve the convergence of chrominance in the overlapping area. By applying the Gaussian pyramid transformation to the edge masks of the two overlapping images respectively, the overlapping weights of the two images are derived, and these weights are blended to generate a high-quality final stitched image. As Figure 1 shown, this step inputs the Laplacian and Gaussian pyramids of image a and image b, and then blends the image weights to generate the final stitched image.
[0046] (8) Operator fusion. The variable memory generated during the operation of the CPU and GPU, as well as the memory access time of these interactions, is extremely large. To address this challenge, inspired by the concept of complex functions in mathematics, the present invention reduces the memory usage and interactive access time and improves the execution speed of the algorithm by integrating the convolutional layer, batch normalization layer, and activation layer.
[0047] Figure 1Shows the overall process from step (1) to step (8).
[0048] As Figure 2 shown, the stitching method of the present invention can be applied to the stitching of fluorescence microscopy images. Figure 2 In (a), there are 25 small field-of-view fluorescence microscopy images obtained by scanning, Figure 2 in (b), there is a large field-of-view, high-resolution fluorescence image obtained by applying the stitching method of the present invention, Figure 2 and in (b1), there is a local enlarged view of the large field-of-view, high-resolution fluorescence image obtained by applying the stitching method of the present invention.
[0049] In addition to fluorescence image stitching, the present invention is still applicable to bright-field images and other modes of microscopy images.
[0050] As Figure 3 shown, the stitching method of the present invention can be applied to the stitching of bright-field microscopy images. Figure 3 In (a), there are 25 small field-of-view bright-field microscopy images obtained by scanning, Figure 3 in (b), there is a large field-of-view, high-resolution bright-field image obtained by applying the stitching method of the present invention, Figure 3 and in (b1), there is a local enlarged view of the large field-of-view, high-resolution bright-field image obtained by applying the stitching method of the present invention.
[0051] Aiming at the problem of the limited field of view of the microscope objective lens, the present invention realizes the effective stitching of the microscopy field of view through a series of innovative technical means. Specifically, the present invention efficiently obtains a series of small field-of-view images by using a microscopy objective lens and an image sensor, and these images cover a large field-of-view range. Adopting a sliding window stitching strategy, combining a zigzag stitching pattern and a multi-dimensional matching technology, improves the calculation accuracy of the stitching position and effectively reduces error accumulation. The present invention introduces a specially designed neural network structure for feature extraction, which includes an input convolutional block, a residual convolutional block and an upsampling layer. After training with a large-scale dataset, it can efficiently and accurately extract key features in microscopy images; the present invention uses a global feature fusion technology based on Transformer and two-dimensional position encoding to improve the matching accuracy of fuzzy regions; finally, through affine transformation, image blending technology of Laplacian pyramid smoothing and operator fusion acceleration technology, seamless image stitching and the improvement of algorithm execution speed are realized. The microscopy field-of-view stitching method of the present invention not only realizes large field-of-view and high spatial resolution microscopy imaging, but also reduces costs, providing important technical support for biologists and medical researchers to observe and analyze samples in a larger range.
[0052] The above are only the preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.
Claims
1. A microscopic field-of-view stitching method with a large field of view, high spatial resolution, and low cost, characterized in that The microscopic field-of-view stitching method includes the following steps: S1. Obtain a series of small-field-of-view images of a sample to be measured covering a large field-of-view range, and construct a small-field-of-view image matrix; S2. For the small-field-of-view image matrix, first stitch the small-field-of-view images in the first row and the first column in sequence as the stitching reference; then use the sliding window technique to stitch the remaining small-field-of-view images in a zigzag pattern in turn; during the stitching process, pair the sliding window with each small-field-of-view image to be stitched, and perform multi-dimensional matching by comparing the horizontal and vertical features in the sliding window with the features of the image to be stitched; the sliding window is constructed based on multiple small-field-of-view images containing image information; S3. Based on a neural network, perform feature extraction on two images to be stitched paired with the sliding window after stitching along the third dimension, generate a feature matrix, and output the coarse-grained local features of the images to be stitched; S4. Based on the Transformer module, convert the coarse-grained local features output by the neural network into global features; use two-dimensional position encoding technology, combine self-attention layers and cross-attention layers, and activate through the Softmax function to generate a confidence matrix; S5. Based on the confidence matrix and the extracted global features, use the mutual nearest neighbor algorithm for feature matching; adopt the random sample consensus method to accurately pair the identified features to form a set of feature pairs; S6. Based on the set of feature pairs, align the images to be stitched through affine transformation, calculate the relative displacement between the target microscopic image stitched by multiple images to be stitched and the sliding window; use the image blending technique with Laplacian pyramid smoothing to perform iterative downsampling on the images to be stitched to improve the convergence of chrominance in the overlapping area and generate the final stitched image.
2. The method for stitching microscopic fields of view with a large field of view, high spatial resolution, and low cost according to claim 1, wherein Step S1 further includes: S11. Place the sample to be measured on the translation stage, and adjust the position of the microscopic objective lens so that the sample presents a clear in-focus image on the image sensor; S12. Control the position of the translation stage where the sample is located to change the sample position. Each time the sample position is changed, use the image sensor to take a picture of the corresponding small-field-of-view area of the sample, scan the sample, and obtain a series of small-field-of-view images covering a large field-of-view range.
3. The microscopic field-of-view stitching method with a large field of view, high spatial resolution, and low cost according to claim 1, wherein In step S2, the process of constructing the sliding window is as follows: Select three adjacent images in the image to be stitched and add a zero-value image for combination; then, based on the center of the combined image, crop an area with the same size as the image to be stitched as the sliding window image.
4. The microscopic field-of-view stitching method with a large field of view, high spatial resolution, and low cost according to claim 1, wherein In step S3, the neural network includes an input convolutional block, a first residual convolutional block, a second residual convolutional block, a third residual convolutional block, a first upsampling layer, and a second upsampling layer connected in sequence; The input convolutional block includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence; the kernel of the convolutional layer of the convolutional layer is 7×7, the stride is 2, and the padding is 3; the input convolutional block performs preliminary feature extraction on the stitched image; The first residual convolution block, the second residual convolution block, and the third residual convolution block each include two convolutional layers, two batch normalization layers, and two ReLU activation layers. Among them, the convolutional layers of the first residual convolution block, the second residual convolution block, and the third residual convolution block all use 3×3 kernels with a padding of 1 and strides of 1, 2, and 3 respectively; the first residual convolution block, the second residual convolution block, and the third residual convolution block further extract the features extracted by the input convolution block, and the extracted features are then subjected to two-step upsampling through the first upsampling layer and the second upsampling layer in sequence to output the key features of the microscopic image. The upsampling operation is implemented using a transposed convolutional layer with a kernel size of 1.
5. The microscopic field-of-view stitching method with a large field of view, high spatial resolution, and low cost according to claim 1, characterized in that Step S4 further includes: The Transformer module uses two-dimensional position encoding technology to encode the rough features, and then injects position awareness to improve the matching accuracy of the blurred regions; the position-encoded features sequentially traverse the self-attention layer and the cross-attention layer to enrich feature extraction; among them, the self-attention layer maintains a consistent input, and the cross-attention layer adjusts the dynamic input according to the direction of attention; after the outputs of the self-attention layer and the cross-attention layer are activated using Softmax, the two resulting activation matrices are multiplied to obtain the confidence matrix.
6. The microscopic field-of-view stitching method with a large field of view, high spatial resolution and low cost according to claim 1, characterized in that In step S6, the process of generating the final stitched image is as follows: The Laplacian pyramid technique is used to smooth and iteratively downsample the overlapping regions of the two images to be stitched to accelerate the convergence of the chromaticity consistency in the overlapping regions; then, the Gaussian pyramid transform is applied to the edges of the two overlapping images respectively to calculate their respective overlapping weights; finally, these overlapping weights are mixed to generate a high-quality final stitched image.
7. The microscopic field-of-view stitching method with a large field of view, high spatial resolution and low cost according to claim 1, characterized in that, The small field-of-view images include fluorescence microscopic images and bright-field microscopic images.