An imaging lens and an image processing method

Abstract

The application discloses an imaging lens and an image processing method, and belongs to the optical technical field. The imaging lens comprises a zoom objective lens; the zoom objective lens comprises a front fixed group, a zoom group, a compensation group, a rear fixed group and an image plane which are coaxially arranged in sequence from an object side to an image side. The application adopts a positive group compensation, optimizes the zoom system by using Zemax, realizes 10 times continuous zooming of 20mm-200mm only through 7 standard spherical lenses, can realize a larger zoom ratio, has a large focal length range, has good imaging quality of the lens, the system cam curve is smooth without breakpoints, and the processing cost of the zoom objective lens is effectively reduced; the end-to-end mapping of the network is learned through grouping dense connection, the network is trained in an unsupervised mode, the fusion image obtained is superior to a traditional image fusion method in visual intuitive feeling and objective evaluation indexes, and manual design of the active level measurement and the fusion rule is avoided.

Description

Technical Field

[0001] This invention relates to the field of optical technology, and in particular to an imaging lens and an image processing method. Background Technology

[0002] Polarization is one of the important physical properties of light. Targets on the Earth's surface or in the atmosphere generate specific polarization information determined by their own characteristics when reflecting, scattering, transmitting, and radiating electromagnetic waves. This polarization information can be used to analyze the target's shape, surface roughness, texture, and the physicochemical properties of its materials. Traditional imaging equipment encodes light intensity and wavelength information into brightness and spectrum for imaging, but cannot obtain target polarization information. Polarization imaging technology, as a cutting-edge technology, adds polarization dimension information to traditional imaging. It not only provides the light intensity distribution in two-dimensional space but also obtains the polarization information of the target and background, and has wide applications in environmental, agricultural, biological, and medical fields.

[0003] As the front optical device of a polarization imaging system, the zoom objective lens undertakes the key task of acquiring target light information. At present, in order to reduce the cost of zoom objectives, two main solutions are used: one is to use plastic lenses instead of glass lenses; the other is to use all standard spherical lenses instead of aspherical lenses. As described by Bai Hubing and Miao Li, Design of Large Aperture Long Focal Length Zoom Optical System [J]. Applied Optics, 2018, 39(5):58-63, the entire system of this design adopts a spherical system and does not use special spherical surfaces such as aspherical or diffractive surfaces. This reduces the design and processing difficulty and manufacturing cost of the overall system. The design has more than 9 lenses, the whole system is relatively complex, and the assembly error is large. For example, Wang Haiyan et al. Design of low-cost high-magnification optical zoom mobile phone camera lens [J]. Progress in Laser & Optoelectronics, 2011, 48(12): 5., mainly uses lenses made of optical plastics. Although the processing cost is low, it is sensitive to temperature. When the temperature changes greatly, it will seriously affect the imaging quality of the lens. For objective lenses that work in the infrared band, it is also not suitable due to the thermal effect of infrared light. This greatly limits the working scenarios of zoom lenses, so they cannot be widely used.

[0004] There are currently two main approaches to image fusion: traditional methods and deep learning-based methods. Traditional methods can be further categorized based on their theoretical foundations, including multi-scale transformations, sparse representations, and subspaces. Among these, multi-scale analysis methods are the most commonly used, such as Laplacian pyramid (LP), low-pass ratio pyramid (RP), discrete wavelet transform (DWT), dual-tree complex wavelet transform (DTCWT), guided filtering (GF), multi-resolution singular value decomposition (MSVD), and non-subsampled contour transform (NSCT). However, all of these methods require applying the same image transformation to different source images, which is not optimal for fusing polarized images. Furthermore, the activity level measurement and fusion rule design in most methods are manually executed, making it difficult to adapt to different scenarios. Deep learning-based methods have become a popular approach in recent years, addressing polarization image fusion using an unsupervised deep network (PFNet). The network comprises an encoder, a feature fusion module, and a decoder. Notably, the encoder incorporates dense blocks to extract more features from the source images. Furthermore, the network is trained using a custom multi-scale weighted structural similarity network loss function and the average absolute difference between the source and fused images as the network loss function. This approach achieves relatively ideal fusion results while avoiding the need for manually designing activity level measurements and fusion rules, making deep learning methods more suitable for engineering applications.

[0005] Polarization imaging systems utilize polarization imaging technology to reconstruct target images. The zoom lens, as the front optical component of the polarization imaging system, plays a crucial role in acquiring target light information; therefore, the design of the polarization imaging optical system is extremely important. Polarization imaging systems operate in the near-infrared band and require long-distance imaging. However, near-infrared zoom lenses are relatively scarce and expensive. This is partly because zoom lenses generally use a large number of aspherical elements; extensive research on zoom lenses reveals that most zoom lenses have more than nine elements, leading to high manufacturing and assembly costs. Another reason is the relatively high difficulty in achieving a 10x continuous zoom design.

[0006] Polarized images acquired directly by polarization imaging systems require further processing through image enhancement algorithms to improve image quality. Image fusion is one such method. Traditional image fusion algorithms typically involve three steps: image transformation, activity level measurement, and fusion rule design to fuse two images. Most methods apply the same image transformation to different source images, which is not optimal for fusing polarized images. Furthermore, the activity level measurement and fusion rule design in most methods are manually executed, making it difficult to adapt to different scenarios.

[0007] This invention discloses an imaging lens and an image processing method. The optical lens designed in this invention employs positive group compensation and utilizes Zemax (optical product design and simulation software) to optimize the zoom system. It can achieve 10x continuous zoom from 20mm to 200mm using only 7 standard spherical lenses, offering a wide focal length range, excellent lens image quality, and a smooth, uninterrupted cam curve, effectively reducing the manufacturing cost of zoom lenses in polarization imaging systems. The polarization image fusion algorithm provided in this invention is based on deep learning. It utilizes a grouped densely connected network to learn end-to-end mappings and trains the network in an unsupervised manner. The fused image obtained in this way outperforms traditional image fusion methods in terms of visual perception and various objective evaluation indicators. Furthermore, it avoids the need for manually designing active level measurements and fusion rules, making this method more practical in engineering. Summary of the Invention

[0008] To address the aforementioned issues, this paper provides an imaging lens and image processing method. By employing positive group compensation and utilizing Zemax to optimize the zoom system, a 10x continuous zoom from 20mm to 200mm can be achieved using only 7 standard spherical lenses. This results in a wide focal length range, excellent lens image quality, and a smooth, uninterrupted cam curve, effectively reducing the manufacturing cost of the zoom objective lens in the polarization imaging system. Furthermore, a deep learning-based polarization image fusion algorithm is used. This algorithm employs a grouped dense connection network to learn end-to-end mappings and trains the network in an unsupervised manner. The resulting fused image outperforms traditional image fusion methods in terms of visual perception and various objective evaluation metrics, while also avoiding the need for manually designing activity level measurements and fusion rules.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows.

[0010] An imaging lens includes a zoom objective for acquiring target light; the zoom objective is composed of 7 lenses, each with a standard spherical surface; the zoom objective includes a front fixed group, a zoom group, a compensation group, a rear fixed group, and an image plane arranged coaxially from the object side to the image side; the front fixed group, the compensation group, and the rear fixed group have positive optical power, and the zoom group has negative optical power.

[0011] Preferably, the front fixed group includes a cemented doublet positive lens with its convex surface facing the object; the zoom group includes a meniscus positive lens with its concave surface facing the image and a biconvex negative lens; the compensation group includes two positive lenses, a first positive lens and a second positive lens, both with their convex surfaces facing the object; the rear fixed group includes two positive lenses, a third positive lens and a fourth positive lens, both with their convex surfaces facing the object; the positions of the front fixed group, the rear fixed group, the image plane, and the F-number remain unchanged during zooming.

[0012] Preferably, the distance between the meniscus positive lens and the biconvex negative lens in the zoom group remains constant during the zoom process, and they move together in a zoom motion with a straight line as the moving curve.

[0013] Preferably, the distance between the first positive lens and the second positive lens in the compensation group remains constant during the zoom process.

[0014] Preferably, when the zoom group performs zoom movement, the first positive lens and the second positive lens together perform a compensation movement with a moving curve to keep the position of the object on the image plane unchanged.

[0015] Preferably, the compensation group and the rear fixing group are coaxially provided with an aperture.

[0016] Preferably, the total travel distance of the variable magnification group is 96 mm.

[0017] Preferably, the total travel distance of the compensation group is 34.48 mm.

[0018] Preferably, the imaging lens has a total length of 280mm, a working wavelength of 940nm, a horizontal field of view of 2W = 1.7°-17°, an MTF of >0.3 at 120lp / mm across the entire field of view, an F-number of 3.3, a maximum distortion of <4%, and an effective focal length EFL of 20mm-200mm.

[0019] Preferably, an image processing method includes an encoder and a decoder, comprising the following steps:

[0020] Step 1: The encoder is used to receive the input images S0 and DoLP. It first extracts shallow features through a convolutional layer, and then extracts deep feature maps from the two source images through grouped dense connection blocks (GDB).

[0021] Step 2: The deep feature maps of the two source images are concatted along the depth direction and fused to obtain a fused feature map;

[0022] Step 3: The fused feature map is decoded by the decoder and the fused image is output; the decoder consists of four convolutional layers connected in series.

[0023] Preferably, the Grouped Dense Connectivity Block (GDB) is composed of three cascaded dense blocks, and its output includes the input of each dense block and the output of the cascaded dense blocks.

[0024] Preferably, each output of the dense block is fed back as input to the next layer to retain more useful information and suppress overfitting.

[0025] Preferably, the image processing method uses multi-scale weighted structural similarity (MSWSSIM) and mean absolute error (MAE) as network loss functions to train the network in an unsupervised deep learning manner.

[0026] By adopting the above technical solution, the present invention has the following beneficial effects.

[0027] (1) This invention employs positive group compensation and utilizes the Zemax optimized zoom system to achieve 10x continuous zoom from 20mm to 200mm using only 7 standard spherical lenses. It can achieve a large zoom ratio, a wide focal length range, and good lens imaging quality. The system cam curve is smooth and without breaks, effectively reducing the manufacturing and processing costs of the zoom objective lens for the polarization imaging system. At the same time, it can greatly simplify the manual adjustment and operation process of the zoom group and compensation group during zooming.

[0028] (2) By using a glass lens, the present invention can effectively avoid the problem of plastic lenses deforming due to thermal effects when working in the near-infrared band, thus ensuring the image quality of the lens.

[0029] (3) All seven optical lenses of the present invention adopt a standard spherical design, eliminating the need for custom-made and processed irregular aspherical lenses, which can effectively reduce the processing cost of zoom lenses and improve the imaging quality of the lens.

[0030] (4) The present invention is based on a deep learning method, which uses a grouped dense connection network to learn end-to-end mapping and trains the network in an unsupervised manner. The fused image obtained in this way is superior to the traditional image fusion method in terms of visual perception and various objective evaluation indicators. At the same time, it avoids the need to manually design activity level measurement and fusion rules, making the method more practical in engineering. Attached Figure Description

[0031] The following provides a detailed discussion of the manufacture and application of preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are merely illustrative of specific ways of manufacturing and using the present invention and do not limit the scope of the invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort.

[0032] Figure 1 This is a diagram of the optical system of the zoom objective lens of the present invention at a short focal length.

[0033] Figure 2 This is a diagram of the optical system of the zoom objective lens of the present invention at long focal length.

[0034] Figure 3This is the optical transfer function diagram of the zoom objective lens of the present invention at short focal length (cutoff frequency is 120 lp / mm).

[0035] Figure 4 This is the optical transfer function diagram of the zoom objective lens of the present invention at long focal length (cutoff frequency is 120 lp / mm).

[0036] Figure 5 This is a field curvature distortion diagram of the zoom objective lens of the present invention at a short focal length.

[0037] Figure 6 This is a field curvature distortion diagram of the zoom objective lens of the present invention at long focal length.

[0038] Figure 7 This is a cam curve diagram of the zoom objective lens of the present invention.

[0039] Figure 8 This is a network architecture diagram of the present invention.

[0040] Figure 9 This is a detailed diagram of the Grouped Dense Connectivity Block (GDB) of the present invention.

[0041] Figure 10 This is a detailed diagram of the dense block of the present invention.

[0042] Figure 11 This is a comparison chart of the polarization image fusion results between the algorithm of this invention and the traditional algorithm.

[0043] in, F 1-Front fixed group; F 2-Variable multiplication group; F 3-Compensation Group; F 4 - Rear fixed group; L1 - Cemented doublet positive lens; L2 - Meniscus positive lens; L3 - Biconvex negative lens; L4 - Positive lens one; L5 - Positive lens two; L6 - Positive lens three; L7 - Positive lens four; S - Aperture stop; S1~S15 - Lens surfaces; I - Image plane. Detailed Implementation

[0044] The following provides a detailed discussion of the manufacture and application of preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are merely illustrative of specific ways of manufacturing and using the invention and do not limit the scope of the invention.

[0045] An imaging lens includes a zoom objective lens for acquiring target light; the zoom objective lens is composed of seven lenses, each with a standard spherical surface; the zoom objective lens includes a front fixed group arranged coaxially from the object side to the image side. F 1. Variable magnification group F 2. Compensation Group F3. Rear fixed group F 4 and image plane I; the front fixation group F 1. Compensation Group F 3 and rear fixed group F 4. Positive optical power, the zoom group F 2 has negative optical power. The front fixed assembly... F 1 includes a cemented doublet positive lens L1 with its convex surface facing the object; the zoom group F 2 includes a meniscus lens L with its concave surface facing the image side. 2 and a biconvex negative lens L3; the compensation group F 3 includes two positive lenses, L4 and L5, both with their convex surfaces facing the object; the rear fixing assembly F 4 includes two positive lenses, L6 and L7, with their convex surfaces facing the object; the front fixing assembly F 1. Rear fixed group F 4. The position of image plane I and the F-number remain unchanged during zooming.

[0046] The zoom group F In section 2, the distance between the meniscus positive lens L2 and the biconvex negative lens L3 remains constant during the zoom process, and they move together in a zoom motion with a linear trajectory. The zoom group... F The total travel distance of 2 is 96mm. The compensation group F The distance between the first positive lens L4 and the second positive lens L5 in module 3 remains constant during zooming. The first positive lens L4 and the second positive lens L5 are located in the zoom group. F 2. During zooming, a compensation motion is performed along with the object, using a curved path, to maintain the object's position on image plane I. The compensation group... F The total travel of 3 is 34.48 mm.

[0047] The compensation group F 3 and the aforementioned rear fixation group F A centrally located coaxial aperture S is provided. The imaging lens has a total length of 280mm, a working wavelength of 940nm, a horizontal field of view of 2W = 1.7°-17°, an MTF > 0.3 at 120lp / mm across the entire field of view, an F-number of 3.3, a maximum distortion of <4%, and an effective focal length EFL of 20mm-200mm.

[0048] An image processing method, including an encoder and a decoder, includes the following steps:

[0049] Step 1: The encoder is used to receive the input images S0 and DoLP. It first extracts shallow features through a convolutional layer, and then extracts deep feature maps from the two source images through grouped dense connection blocks (GDB).

[0050] Step 2: The deep feature maps of the two source images are concatted along the depth direction and fused to obtain a fused feature map;

[0051] Step 3: The fused feature map is decoded by the decoder and the fused image is output; the decoder consists of four convolutional layers connected in series.

[0052] The Grouped Dense Connectivity Block (GDB) consists of three cascaded dense blocks, and its output includes the input of each dense block and the output of the cascaded dense blocks. Each output of a dense block is fed back as input to the next layer to retain more useful information and suppress overfitting. The image processing method described uses Multi-Scale Weighted Structural Similarity (MSWSSIM) and Mean Absolute Error (MAE) as network loss functions to train the network in an unsupervised deep learning manner.

[0053] The following is a detailed description with reference to the accompanying drawings in the instruction manual.

[0054] An imaging lens includes a zoom objective for acquiring target light, the optical system diagrams of which are shown below for short focal length and long focal length. Figure 1 , 2 As shown. The zoom objective lens adopts a mechanically compensated positive group compensation method, and the zoom objective lens includes a front fixed group arranged coaxially from the object side to the image side. F 1. Variable magnification group F 2. Compensation Group F 3. Rear fixed group F 4 and image plane I; anterior fixation group F 1. Positive optical power, comprising a cemented doublet positive lens L1 with a convex surface facing the object side, and a spherical surface. The zoom group... F 2 has negative optical power, including a meniscus positive lens L2 with its concave surface facing the image side and a biconvex negative lens L3, both of which have spherical surfaces. During the zoom process, the distance between the meniscus positive lens L2 and the biconvex negative lens L3 remains unchanged, and they move together to achieve the zoom effect. The movement curve is a straight line, and the total movement distance is 96mm.

[0055] The compensation group F 3. Positive optical power, comprising two positive lenses, L4 and L5, with convex surfaces facing the object side. The distance between the positive lenses L4 and L5 remains constant, and both have spherical surfaces. When the zoom group... F When the two lenses move, the compensation group FIn section 3, the first positive lens L4 and the second positive lens L5 move together to maintain the position of image plane I. The movement is curved, with a total travel distance of 34.48 mm. The aperture stop S is located in the compensation group. F 3 and rear fixed group F Between 4, the rear fixing group F 4. It has positive optical power and includes two positive lenses, L6 and L7, with their convex surfaces facing the object side. Both of these lenses have spherical surfaces. During zooming, the cemented doublet positive lens L1, L6, L7, their F-number, and the image plane I all remain unchanged.

[0056] The zoom objective lens maintains a constant F-number (a parameter of the lens's light transmission capability) during zooming, and the parameters of this invention during continuous zooming are set as follows: effective focal length EFL = 20mm-200mm, F-number = 3.3, total optical system length = 280mm, working wavelength = 940nm, horizontal field of view range 2W = 1.7°-17°, MTF > 0.3 at 120lp / mm across the entire field of view, maximum distortion < 4%, and total number of lenses = 7.

[0057] The zoom lens designed in this invention employs positive group compensation and utilizes Zemax (optical design and simulation software) to optimize the zoom system. This zoom lens uses 7 optical elements, significantly reducing lens production and manufacturing costs compared to zoom lenses on the market that generally have more than 9 elements. It also greatly simplifies manual adjustment of the zoom group during zooming. F 2 and Compensation Group F 3. Modulation and Operation Process: All seven optical lenses of this invention employ a standard spherical surface design, eliminating the need for custom-made and manufactured irregular aspherical lenses. This effectively reduces the processing cost of the zoom lens and improves its imaging quality. This invention provides a 10x continuous zoom range of 20mm-200mm using seven standard spherical lenses, offering a wide focal length range and excellent image quality. Furthermore, the use of glass lenses effectively avoids the thermal deformation of plastic lenses during near-infrared operation, which can affect the imaging quality of the zoom lens and ensure its quality.

[0058] The actual changes in the optical transfer function curves of the zoom lens of this invention at short and long focal lengths are as follows: Figure 3 , 4 As shown, at a spatial frequency of 120 lp / mm, the modulation transfer function (MTF) of the full-field imaging lens is greater than 0.3, close to the diffraction limit, indicating that the imaging lens of this invention has a strong ability to reproduce the image of the real object, and the imaging lens transmits optical images with less distortion; the optical transfer function curve in the figure is relatively flat in both short-crossing and long-focal-length states, indicating that the imaging lens has good consistency between the edges and the center, and has good image quality.

[0059] The field distortion of the zoom lens of this invention at short and long focal lengths is as follows: Figure 5 , 6 As shown, the zoom objective lens of this invention exhibits a field area of ​​less than 0.1 mm at both short focal length (F=20mm) and long focal length (F=200mm), while the maximum distortion is less than 4%. The low degree of distortion and deformation meets the measurement and imaging requirements of polarization imaging systems for field area and distortion, demonstrating excellent imaging performance.

[0060] The cam curve diagram of the zoom objective lens of this invention is as follows: Figure 7 As shown, it can be seen that during the zooming process, the zoom group of the zoom objective lens of the present invention... F 2. Perform linear motion, compensation group F 3. Perform curved motion to ensure that the object's position on image plane I remains unchanged throughout the focusing process. F 2 and the compensation group F The smooth cam curve of 3 has no breaks or inflection points, indicating that the zoom objective lens of the present invention does not drift at any focal length, and has high stability.

[0061] The network architecture diagram of this invention is as follows: Figure 8 As shown in the diagram. S0 and DoLP represent the input image, C1, C11, C12, C13, and C14 represent convolutional layers, R1, R11, R12, R13, and R14 represent ReLU activation functions, GDB represents grouped densely connected blocks, F1 and F2 represent the extracted feature maps, Concat2 represents the connection layer, F12 represents the fused feature map, and F represents the fused image. The arrows in the diagram indicate downward propagation operations. This invention employs the following image processing method for image fusion, including an encoder and a decoder, with the specific steps as follows.

[0062] Step 1: The encoder is used to receive the input images S0 and DoLP. It first extracts shallow features through a convolutional layer, and then extracts deep feature maps from the two source images through grouped dense connection blocks (GDB).

[0063] Step 2: The deep feature maps of the two source images are concatted along the depth direction and fused to obtain a fused feature map;

[0064] Step 3: The fused feature map is decoded by the decoder and the fused image is output; the decoder consists of four convolutional layers connected in series.

[0065] Detailed diagrams of the Grouped Dense Connectivity Block (GDB) of the present invention are shown below. Figure 9As shown in the diagram, the Grouped Dense Block (GDB) is composed of three cascaded dense blocks. Its output includes the input of each dense block and the output of the cascaded dense blocks. In the diagram, DB1, DB2, and DB3 represent dense blocks, Concat1 represents the connection layer, thick arrows indicate downward propagation operations, and thin arrows represent the outputs of DB1, DB2, and DB3. Concat1 connects the outputs of the thick and thin arrows along the depth direction.

[0066] Detailed diagrams of the dense block of the present invention are shown below. Figure 10 As shown in the diagram, each output of the dense block is fed back as input to the next layer to retain more useful information and suppress overfitting. Taking DB1 as an example, the remaining DB2 and DB3 have the same structure. In the diagram, C2, C3, and C4 represent convolutional layers, R2, R3, and R4 represent ReLU activation functions, thick arrows indicate downward propagation operations, and thin arrows indicate the corresponding outputs. Similarly, DB2 contains C5, C6, C7, R5, R6, and R7; DB3 contains C8, C9, C10, R8, R9, and R10.

[0067] The comparison of polarization image fusion results between the algorithm of this invention and the traditional algorithm is shown in the figure below. Figure 11 As shown in the figure, the algorithm of this invention produces images with higher brightness, richer details, and better conformity to human visual perception compared to other methods.

[0068] Table 1 shows the optical structural parameters of the zoom objective lens of the present invention at long focal length.

[0069]

[0070] Among them, S1~S15 are lens surfaces. All lens surfaces in this invention are standard spherical surfaces. That is, all 15 surfaces of the 7 optical lenses in this invention adopt a standard spherical design. There is no need to customize and process irregular aspherical lenses, which can effectively reduce the processing cost of zoom lenses and improve the imaging quality of the lens.

[0071] Table 1 is also combined with Figure 1 , 2It can be seen that the lens surfaces of the cemented doublet positive lens L1 are S1, S2, and S3; the lens surfaces of the meniscus positive lens L2 are S4 and S5; the lens surfaces of the biconvex negative lens L3 are S6 and S7; the lens surfaces of the first positive lens L4 are S8 and S9; the lens surfaces of the second positive lens L5 are S10 and S11; the lens surfaces of the third positive lens L6 are S12 and S13; and the lens surfaces of the third positive lens L7 are S14 and S15. It is evident that this invention uses only 7 lenses, and the radius of curvature of each lens is shown in Table 1. Compared to zoom lenses on the market that generally have more than 9 lenses, this invention significantly reduces the production and manufacturing costs of the lens by increasing the number of optical lenses to 7, while also greatly simplifying the manual adjustment of the zoom group during zooming. F 2 and Compensation Group F The modulation and operation process of 3 is simpler.

[0072] Furthermore, as shown in Table 1, the surfaces of all seven optical lenses in this invention are designed with standard spherical surfaces, eliminating the need for custom-made and processed irregular aspherical lenses. This effectively reduces the processing cost of zoom lenses and improves the imaging quality of the lens.

[0073] Table 2 shows the parameter settings for the polarization image fusion network of this invention.

[0074]

[0075] In this context, Common layer i (i=1, 2, 3, 4, 5) represents a common convolutional layer, and DB j (j=1, 2, 3) represents three cascaded dense blocks in the Grouped Dense Block (GDB). The GDB is composed of three cascaded dense blocks, and its output includes the input of each dense block and the output of the cascaded dense blocks. All convolutional layers are 3×3 in size, with a stride of 1, and all use the same padding. The encoder has 16 output channels and uses ReLU activation. The decoder's convolutional layers have progressively fewer output channels: 64, 32, 16, and 1, and all except the last layer use ReLU activation.

[0076] Table 3 shows a comparison of the quantitative evaluation indicators between the algorithm of this invention and the traditional algorithm.

[0077]

[0078] Among them, entropy (EN), mutual information (MI), standard deviation (SD), visual information fidelity (VIF), sum of differential correlations (SCD), edge similarity measure (Qabf), structural similarity index measure (SSIM), and multi-scale structural similarity index measure (MS_SSIM) are all evaluated, with higher values ​​being better. To enable quantitative comparison between the proposed fusion method and other existing methods, eight fusion image quality evaluation metrics were used. As shown in Table 3, all metrics of the present invention achieved the highest values ​​for the eight fusion image quality metrics, demonstrating that the algorithm of the present invention is significantly superior to 11 traditional algorithms such as Bayesian, CWT, and DTCWT.

[0079] The image processing method described in this invention uses multi-scale weighted structural similarity (MSWSSIM) and mean absolute error (MAE) as network loss functions to train the network in an unsupervised deep learning manner. The specific network loss function used in this invention is as follows.

[0080] Network loss function L It consists of two parts:

[0081]

[0082] in, L MSWSSIM For multi-scale weighted structural similarity loss, L P For pixel loss, l The weights are used to balance the two losses.

[0083] Multi-scale weighted structural loss function L MSWSSIM for:

[0084]

[0085] in, c ω Indicates the weighting coefficient. L SSIM ( x , y ; oh )express x and y With window oh The local similarity of can be expressed as:

[0086]

[0087] in, oh x It is an image x In the window oh The area within, yes oh x The average value, and They are respectively oh x variance and oh x and oh y Covariance, similarly, oh y , , It means the same thing. C 1 and C 2 is a constant. L MSWSSIM Using multiple windows to extract features at different scales is more efficient than using only one window. SSIM It is more accurate as a loss function.

[0088] Weighting coefficient c ω The definition of is:

[0089]

[0090] in, g ( x )=max( x ,0.0001), max(a,b) represents the selection a and b The larger ones.

[0091] Pixel loss L P for:

[0092]

[0093] in, M × N Indicates the size of the image. The L1 norm is represented by the formula: .

[0094] The unsupervised deep learning approach used in this invention for network training is configured as follows. Deep learning algorithms require network training. A total of 240 polarization images were acquired using a polarization camera, with 150 used for training, 50 for validation, and 40 for testing. Each image is 480×640 pixels. During training, the images were cropped into 40×40 blocks, which were then flipped and rotated, resulting in over 5 million blocks for data augmentation. The Adam optimization function was used, with batch size set to 128, epochs set to 20, and the loss function set to 0.1 and 1×10⁻⁶.-4 and 9×10 -4 .

[0095] Although the specification has provided a detailed description, it should be understood that various changes, substitutions, and modifications can be made without departing from the spirit and scope of the invention as defined by the appended claims. Furthermore, the specific embodiments described are not intended to limit the scope of the invention, and those skilled in the art will readily understand based on this invention that existing or future-developed processes, machines, manufactures, compositions of matter, means, methods, or steps can perform substantially the same functions or achieve substantially the same results as the embodiments of the invention. Therefore, the appended claims are intended to include such processes, machines, manufactures, compositions of matter, means, methods, or steps within their scope.

Claims

1. An imaging lens, characterized in that: Includes a zoom objective lens; the zoom objective lens is composed of 7 lenses, each with a standard spherical surface; the zoom objective lens includes a front fixed group, a zoom group, a compensation group, a rear fixed group, and an image plane arranged coaxially from the object side to the image side; the front fixed group, the compensation group, and the rear fixed group have positive optical power, and the zoom group has negative optical power; The front fixed group includes a cemented doublet positive lens with its convex surface facing the object; the zoom group includes a meniscus positive lens with its concave surface facing the image and a biconcave negative lens; the compensation group includes two positive lenses, positive lens one and positive lens two, both with their convex surfaces facing the object; the rear fixed group includes positive lens three and positive lens four; the position of the front fixed group, the rear fixed group, the image plane, and the F-number remain unchanged during zooming.

2. An imaging lens as described in claim 1, characterized in that: The distance between the meniscus positive lens and the biconcave negative lens in the zoom group remains constant during the zoom process, and they move together in a linear zoom motion.

3. An imaging lens as described in claim 2, characterized in that: The distance between positive lens one and positive lens two in the compensation group remains unchanged during the zoom process; when the zoom group performs zoom movement, positive lens one and positive lens two together perform a compensation movement with a curved trajectory to keep the position of the object on the image plane unchanged.

4. An imaging lens as described in claim 3, characterized in that: The compensation group and the rear fixing group are coaxially provided with an aperture in the middle.

5. An imaging lens as described in claim 1, characterized in that: The total travel distance of the zoom group is 96 mm; the total travel distance of the compensation group is 34.48 mm.

6. The imaging lens as described in any one of claims 1-5, characterized in that: The imaging lens has a total length of 280mm, a working wavelength of 940nm, a horizontal field of view of 2W = 1.7°-17°, an MTF of >0.3 at 120lp / mm across the entire field of view, an F-number of 3.3, a maximum distortion of <4%, and an effective focal length EFL of 20mm-200mm.

Images