Iris and pupil real-time segmentation method based on deformable convolution and ellipse prior constraint

Abstract

The invention discloses an iris pupil real-time segmentation method based on deformable convolution and ellipse prior constraint, and relates to the technical field of computer vision and biological feature recognition. The method comprises the following steps: acquiring an original iris image, and preprocessing the original iris image by adopting dynamic shielding enhancement; constructing a lightweight encoder-decoder network, wherein the lightweight encoder-decoder network comprises an encoder, a decoder and a post-processing module; the encoder adopts a cascaded deformable convolution module and an elliptical attention mechanism module to perform feature extraction; training and optimizing the lightweight encoder-decoder network; carrying out multi-task loss calculation, introducing an elliptical parameter regression constraint design loss function on the basis of segmentation loss, and optimizing a segmentation result; and ellipse parameter refinement is carried out through a post-processing module, and ellipse fitting optimization is carried out on a segmentation result. Compared with the prior art, the iris pupil real-time segmentation method has the advantages that the precision is improved, the speed is optimized, and the robustness is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and biometric recognition, and particularly relates to a real-time iris pupil segmentation method based on deformable convolution and elliptical prior constraint. Background Art

[0002] The photoreflectometry method is a pupil examination method that can evaluate the pupil's response to light stimulation. However, this traditional pupil measurement method has obvious limitations. It relies on the naked eye to observe the pupil response after light source stimulation, and usually the measurement error is as high as 20%-40%. These defects seriously affect the measurement accuracy and reliability, and urgently need to be solved through technological innovation. With the development of deep learning technology, the iris pupil segmentation method based on deep learning has become the current mainstream research direction.

[0003] However, the existing iris pupil segmentation methods based on deep learning face three key challenges:

[0004] 1) Shape distortion problem; traditional segmentation networks (such as U-Net) adopt a standard convolution kernel structure and cannot effectively adapt to the special elliptical geometric features of the iris pupil, especially showing serious deficiencies in occlusion scenarios. Experimental data shows that when the upper eyelid occlusion degree exceeds 30%, the IoU index of the traditional method drops by up to 42.7%, resulting in serious deformation of the segmentation result and not conforming to the geometric distribution law of biometric features.

[0005] 2) Computational redundancy problem; the parameter quantity of mainstream semantic segmentation networks (such as DeepLabv3+) generally exceeds 60M, the model volume is large and the computational complexity is high, making it difficult to be efficiently deployed on mobile devices. Test data shows that on the Snapdragon 855 mobile platform, the inference latency of such models reaches 380ms, far exceeding the millisecond-level response requirement for real-time interaction, and restricting the application on low-power devices such as smart glasses.

[0006] 3) Post-processing missing problem; existing methods generally lack a geometric constraint and correction mechanism for the segmentation result. Especially when there are quality problems such as motion blur in the image, the edge sawtooth phenomenon of the segmentation result is serious. Measured data shows that the edge smoothness drops by 58.3%, affecting the stability of subsequent feature extraction.

[0007] In summary, aiming at the limitations of the research on iris pupil segmentation based on deep learning, the present invention proposes a three-dimensional iris real-time segmentation scheme for mobile terminals, innovatively proposes a deformable elliptical convolution operator and a geometric constraint attention mechanism. By dynamically adjusting the receptive field of the convolution kernel to adapt to the iris physiological characteristics and combining parametric post-processing to correct the segmentation result, the model is compressed to 2.1MB while ensuring an accuracy of more than 98%. It is particularly suitable for the live detection scenario of low-power devices such as smart glasses. Summary of the Invention

[0008] The purpose of the present invention is to provide a real-time iris pupil segmentation method based on deformable convolution and elliptical prior constraint, so as to solve the problems of shape distortion, computational redundancy and post-processing deficiency existing in the prior art of real-time iris pupil segmentation method proposed in the above background technology.

[0009] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0010] In the first aspect, the present invention proposes a real-time iris pupil segmentation method based on deformable convolution and elliptical prior constraint, including the following steps:

[0011] S1. Collect the original iris image; preprocess the original iris image by using dynamic occlusion enhancement;

[0012] S2. Construct a lightweight encoder-decoder network; the lightweight encoder-decoder network includes an encoder, a decoder and a post-processing module;

[0013] The encoder adopts a cascaded deformable convolution module and an elliptical attention mechanism module for feature extraction; the deformable convolution module enhances the ability to extract pupil edge features through an adaptive deformable convolution kernel; the elliptical attention mechanism module uses the iris anatomical prior to generate an elliptical spatial constraint for matching the attention mask and the biometric distribution;

[0014] S3. Train and optimize the lightweight encoder-decoder network; perform multi-task loss calculation, introduce an elliptical parameter regression constraint on the basis of the segmentation loss to design a loss function, and optimize the segmentation result;

[0015] S4. Perform elliptical parameter refinement through the post-processing module, and perform elliptical fitting optimization on the segmentation result.

[0016] Preferably, the preprocessing by using dynamic occlusion enhancement is specifically to preprocess the original iris image through a combination strategy of random region erasing and elastic deformation.

[0017] Preferably, the encoder performs first-level feature extraction, elliptical attention enhancement and second-level downsampling, and the decoder performs upsampling, skip connection, segmentation head output and elliptical parameter prediction.

[0018] Preferably, the deformable convolution module adopts a method of adapting the deformable convolution kernel to the iris edge curvature, and the mathematical expression of the deformable convolution kernel is specifically as follows:

[0019] x l+1 = ReLU(BN(DefromConv(DefromConv(x l ))))

[0020]

[0021]

[0022] Among them, p represents the spatial position coordinates of the output feature map; K is the sampling points of the 3×3 convolution kernel; Δp k ∈R 2 is the learnable offset; f offset is the offset prediction function;

[0023] Through the dynamic adjustment of Δp k the convolution kernel can adapt to the curvature features of the iris texture.

[0024] Preferably, the elliptical attention mechanism module generates an elliptical heat map through elliptical parameter prediction, specifically as follows:

[0025] First, parameterize the ellipse as follows:

[0026] Θ = {c x ∈[0,1], c y ∈[0,1], a∈[0, W / 2], b∈[0, H / 2], θ∈[-π, π]}

[0027] The parameter prediction process is as follows:

[0028]

[0029]

[0030]

[0031] Among them, F avy is the feature vector after global average pooling;

[0032] Secondly, generate the elliptical heat map;

[0033] The elliptical membership function is as follows:

[0034]

[0035] The rotation coordinate transformation is as follows:

[0036] x′ θ = (x - c x W)cosθ + (y - c y H)sinθ

[0037] y ′ θ = -(x - c x W)sinθ + (y - c yH) cosθ

[0038] Finally, feature enhancement is performed by fusing the elliptical heatmap with the original features:

[0039]

[0040] where ⊙ represents element-wise multiplication.

[0041] Preferably, the loss function adopts a dual-branch decoder combined with a joint loss function: the segmentation head outputs pixel-level probabilities:

[0042]

[0043] The elliptical parameter regression head predicts iris and pupil parameters:

[0044]

[0045] The joint loss function is:

[0046]

[0047] where λ1, λ2, λ3, and λ4 represent balance coefficients.

[0048] Furthermore, the joint loss function is specifically:

[0049] Cross-entropy loss:

[0050]

[0051] Dice loss:

[0052]

[0053] Elliptical parameter regression loss:

[0054]

[0055] Shape consistency loss:

[0056]

[0057] where represents the cross-entropy loss, represents the Dice loss, represents the elliptical parameter regression loss, represents the shape consistency loss.

[0058] Preferably, the post-processing module adopts an elliptical constraint optimization function, specifically as follows:

[0059]

[0060] Among them, Ψ k represents the extraction of iris pupil ellipse parameters, is the Gaussian prior distribution, and ∑ is the credibility matrix of the prediction parameters.

[0061] In a second aspect, the present invention proposes an iris pupil real-time segmentation system, including:

[0062] A lightweight encoder, including a deformable convolution module and an ellipse attention mechanism module, for feature extraction;

[0063] A decoder for multi-scale feature fusion, including a convolutional layer, an ellipse parameter prediction head, and a segmentation head, for multi-scale feature fusion;

[0064] An ellipse parameter-guided post-processing module for optimizing the iris segmentation prediction result.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] (1). The present invention constructs a lightweight encoder-decoder network containing a deformable convolution module, enhances the ability to extract pupil edge features through an adaptive deformable convolution kernel; designs an ellipse attention mechanism module, generates an ellipse space constraint using the iris anatomical prior, and realizes the matching of the attention mask and the biometric feature distribution; innovatively adopts a shape-aware joint loss function, introduces an ellipse parameter regression constraint on the basis of the traditional segmentation loss, and improves the morphological consistency by synchronously optimizing the segmentation result and the ellipse fitting accuracy. The system integrates dynamic occlusion enhancement preprocessing, adopts a combination strategy of random region erasing and elastic deformation to improve the robustness of the model, and cooperates with the ellipse fitting correction module in the post-processing stage to effectively solve the problem of the decrease in segmentation accuracy of traditional methods in cases of occlusion, deformation, etc. while maintaining high real-time performance. This solution realizes end-to-end accurate biometric segmentation through the collaborative optimization of deformable convolution and geometric constraints.

[0067] (2). The iris pupil real-time segmentation method in the present invention has improved accuracy compared with the prior art. The iris pupil real-time segmentation method in the present invention reaches 98.2% IoU on the CASIA-IrisV4 dataset, an improvement of 9.3% compared with the baseline model.

[0068] (3). The iris pupil real-time segmentation method in the present invention has optimized speed compared with the prior art. The number of model parameters in the present invention is only 2.1M, and the inference time on the iPhone 13 is 23ms.

[0069] (4). The iris pupil real-time segmentation method in the present invention has strong robustness compared with the prior art. The iris pupil real-time segmentation method in the present invention still maintains a recognition rate of 91.5% under 30% occlusion. Description of the Drawings

[0070] Figure 1 This is the flowchart of the real-time iris pupil segmentation method based on deformable convolution and elliptical prior constraint in the present invention;

[0071] Figure 2 This is the structural block diagram of the lightweight encoder-decoder network in the present invention;

[0072] Figure 3 This is the network optimization schematic diagram of the lightweight encoder-decoder network in the present invention;

[0073] Figure 4 This is the schematic diagram of the elliptical spatial attention mechanism in the present invention;

[0074] Figure 5 This is the flowchart of the shape constraint loss calculation in the present invention;

[0075] Figure 6 This is the comparison chart of the post-processing elliptical fitting effect in the present invention. Detailed implementation manners

[0076] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0077] Embodiment 1:

[0078] Refer to Figure 1 , for the real-time iris pupil segmentation method based on deformable convolution and elliptical prior constraint, the core of this method is to construct a lightweight encoder-decoder network to achieve accurate and efficient iris segmentation. The lightweight encoder-decoder network includes a lightweight encoder, a decoder for multi-scale feature fusion, and a post-processing module guided by elliptical parameters, as Figure 2 shown. The method is as follows:

[0079] Step 1: Obtain the original image and perform image preprocessing.

[0080] S1.1 Image acquisition: Obtain the original image containing the eye region from the imaging device, ensuring that the image resolution meets the segmentation requirements (it is recommended to be at least 640×480 pixels).

[0081] S1.2 Image preprocessing:

[0082] First, apply the Random Erasing technique;

[0083] Randomly select a rectangular area in the image, set the pixel values of the selected area to random values or preset values, and simulate natural occlusion situations such as eyelids, eyelashes, and glasses.

[0084] Secondly, implement the Elastic Deformation strategy;

[0085] Generate a random displacement field, apply a non - linear transformation to the original image, and simulate real - world scenarios such as eyeball rotation and perspective change.

[0086] Finally, perform normalization on the pre - processed image;

[0087] Normalize the pixel values to the range [0, 1], and apply Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance edge features.

[0088] When pre - processing the data, the code for randomly generating an eyelid simulation mask is as follows:

[0089]

[0090] Step 2: Design and implementation of a lightweight network structure; construct a lightweight backbone network, and the lightweight backbone network includes an encoder and a decoder.

[0091] For the encoder - decoder architecture, the specific feature fusion formula is as follows:

[0092]

[0093] Among them, represents the up - sampling operation, represents the channel - compression convolution, represents the element - wise addition.

[0094] S2.1 Encoder design;

[0095] Add an integrated deformable convolution module to the encoder, add a deformable convolution layer to the key layer, design an offset prediction network to adaptively learn deformation parameters, and dynamically adjust the receptive field of the convolution kernel to fit the elliptical shape of the iris and pupil.

[0096] Deformable convolution module: Design a deformable elliptical convolution operator. By dynamically adjusting the receptive field of the convolution kernel, it can adaptively fit the unique physiological characteristics and morphological changes of the iris and pupil, and effectively solve the problem of shape distortion of standard convolution in occlusion scenarios. The specific implementation is to construct a multi - level deformable residual module, and each level contains two deformable convolution layers.

[0097] In this embodiment, the deformable convolution module adopts a two - stage cascaded structure. The first layer generates offsets, and the second layer performs deformable convolution.

[0098] By dynamically learning the offset to adjust the sampling position of the convolutional kernel, the mathematical expression of the deformable convolutional kernel is as follows:

[0099] x l+1 = ReLU(BN(DefromConv(DefromConv(x l ))))

[0100]

[0101]

[0102] where p represents the spatial position coordinates of the output feature map; K is the number of sampling points of the 3×3 convolutional kernel; Δp k ∈R 2 is the learnable offset; f offset is the offset prediction function;

[0103] Through the dynamic adjustment of Δp k the convolutional kernel can adapt to the curvature features of the iris texture.

[0104] The actual convolution operation uses bilinear interpolation to implement deformation sampling.

[0105] An elliptical attention mechanism module is added to the encoder. Based on iris anatomy knowledge, an elliptical spatial prior is generated, and the elliptical prior is converted into an attention mask to enhance the features in the pupil-iris boundary region through the spatial attention mechanism.

[0106] The elliptical attention mechanism module, as Figure 4 shown, is specifically implemented as follows:

[0107] 1) Parameter prediction: Generate elliptical parameters through global pooling and 1×1 convolution.

[0108] First, parameterize the ellipse as follows:

[0109] Θ = {c x ∈[0,1], c y ∈[0,1], a∈[0, W / 2], b∈[0, H / 2], θ∈[-π, π]}

[0110] The parameter prediction process is as follows:

[0111]

[0112] where F avy is the feature vector after global average pooling;

[0113]

[0114] 2) Thermal map generation: Define the elliptical membership function in the rotated coordinate system:

[0115] The rotation coordinate transformation is as follows:

[0116] x ′ θ = (x - c x W) cosθ + (y - c y H) sinθ

[0117] y ′ θ = -(x - c x W) sinθ + (y - c y H) cosθ

[0118] The elliptical membership function is as follows:

[0119]

[0120] Finally, perform feature enhancement and fuse the elliptical thermal map with the original features:

[0121]

[0122] where ⊙ is the element-wise multiplication.

[0123] S2.2 Decoder design;

[0124] Construct a feature decoding and upsampling path, and use transposed convolution or bilinear interpolation upsampling method to achieve skip connection to fuse high and low level features, and gradually restore the spatial resolution to the input size.

[0125] Step 3: Joint optimization training process.

[0126] Train the lightweight encoder-decoder network and optimize the network parameters. Perform dynamic occlusion enhancement on the original image, preprocess the original image by randomly generating eyelid simulation masks as the training set of the lightweight encoder-decoder network, train the lightweight encoder-decoder network, and calculate the loss function based on the prediction results to optimize the lightweight network, as Figure 3 shown.

[0127] Design the loss function, use the dual-branch decoder combined with the joint loss function: automatically optimize the segmentation edge, solve the problem of poor segmentation effect of traditional methods when the image is severely occluded, and make the segmentation edge smoother, as Figure 5 shown.

[0128] The segmentation head outputs pixel-level probabilities:

[0129]

[0130] The elliptical parameter regression head predicts iris and pupil parameters:

[0131]

[0132] S3.1 Shape perception combined loss function design;

[0133] S3.1.1 Implement traditional segmentation loss calculation; combine Dice loss and cross-entropy loss to evaluate pixel-level prediction accuracy, and add a boundary awareness term to enhance edge precision.

[0134] S3.1.2 Implement elliptical parameter regression constraints; predict the elliptical parameters (center coordinates, major and minor axes, rotation angle) of the iris and pupil, calculate the parameter distance between the predicted ellipse and the true annotated ellipse, and use the geometric parameter error as an additional constraint term.

[0135] Combined to get:

[0136]

[0137] Cross-entropy loss:

[0138]

[0139] Dice loss:

[0140]

[0141] Elliptical parameter regression loss:

[0142]

[0143] Shape consistency loss:

[0144]

[0145] S3.2 Multi-task training strategy;

[0146] S3.2.1 Design a weight adaptive adjustment mechanism; dynamically adjust the weights of the segmentation task and the elliptical parameter regression task according to the training progress, focusing on segmentation accuracy in the initial stage of training and strengthening geometric constraints in the later stage.

[0147] S3.2.2 Implement gradient clipping and learning rate scheduling, set a gradient threshold to prevent training instability, and use the cosine annealing learning rate strategy to optimize the convergence process.

[0148] Specifically:

[0149] The first stage: Only train the segmentation backbone (learning rate 1e-3);

[0150] The second stage: Jointly train the parameter prediction head (learning rate 5e-4);

[0151] The third stage: Freeze the encoder and fine-tune the attention module (learning rate 1e-4).

[0152] Step 4: Post-processing and result optimization.

[0153] Ellipse fitting correction is a key step in the post-processing of iris pupil segmentation. It combines biological anatomy knowledge with computer vision technology to solve the shape distortion problem in the initial segmentation result, as Figure 6 shown. The specific steps are as follows:

[0154] S4.1 Edge extraction of the segmentation result;

[0155] Extract the binary masks of the iris region and the pupil region respectively; convert the multi-class segmentation result into separate binary images, with the iris region corresponding to the value 1 and the pupil region corresponding to the value 2.

[0156] Apply the contour extraction algorithm to obtain the edge point set; use the findContours function to identify the edge contours, select the contour with the largest area as the boundary of the target region, and filter out the stray contours with too small area to improve robustness.

[0157] S4.2 Ellipse parameter fitting and optimization;

[0158] Apply the RANSAC ellipse fitting algorithm; perform random sampling on the edge point set, iteratively fit multiple candidate ellipse models, calculate the number of inliers and the fitting error, and select the ellipse model with the most inliers and the smallest error.

[0159] Apply biological anatomy constraints to filter out abnormal results:

[0160] Iris shape constraint: The ratio of the long axis to the short axis is usually between 0.8 and 1.2;

[0161] Pupil shape constraint: The ratio range is more strict, usually between 0.85 and 1.15

[0162] Position constraint: The pupil center must be located inside the iris

[0163] Size constraint: The pupil diameter is usually 15%-60% of the iris diameter

[0164] S4.3 Reconstruct and optimize the segmentation mask;

[0165] Generate a regular shape based on the fitted ellipse parameters; use the ellipse parameters (center coordinates, lengths of the major and minor axes, rotation angle) to draw a standard ellipse and generate separate mask images for the iris and the pupil.

[0166] Apply morphological post-processing; ensure that the iris and pupil boundaries are smooth and continuous, and eliminate the jagged edges and holes in the segmentation result

[0167] Synthesize the final segmentation result; assign the iris region a medium gray value (128) and the pupil region a high gray value (235), and combine them to form the final classification mask.

[0168] The post-processing module adopts an elliptical constraint optimization function, specifically as follows:

[0169]

[0170] Among them, Ψ k represents the extraction of iris-pupil ellipse parameters, is the Gaussian prior distribution, and ∑ is the credibility matrix of the prediction parameters.

[0171] In the present invention, the lightweight network adopts GPU-CPU heterogeneous computing, where TensorRT is used to accelerate the deformable convolution calculation, and the ellipse fitting post-processing is executed in parallel on the CPU.

[0172] The code for the overall model quantization is as follows:

[0173] quant_model = torch.quantization.quantize_dynamic(

[0174] model, {nn.Conv2d, nn.Linear}, dtype = torch.qint8)

[0175] In the present invention, the concise and innovative network structure compresses the model volume to 0.5MB while maintaining a segmentation accuracy of more than 92%, and is particularly suitable for real-time live detection scenarios of low-power mobile devices such as smart glasses.

[0176] The above is only used to help understand the method and its core essence of the present invention, but the protection scope of the present invention is not limited thereto. For those of ordinary skill in the art in the technical field of the present invention, any equivalent replacement or change made according to the technical solution and inventive concept of the present invention should be covered within the protection scope of the present invention. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A real-time iris and pupil segmentation method based on deformable convolution and elliptical prior constraint, characterized in that It includes the following steps: S1. Collect the original iris image; perform preprocessing on the original iris image using dynamic occlusion enhancement; S2. Construct a lightweight encoder-decoder network; the lightweight encoder-decoder network includes an encoder, a decoder, and a post-processing module; The encoder uses a cascaded deformable convolution module and an elliptical attention mechanism module for feature extraction; The deformable convolution module enhances the ability to extract pupil edge features through an adaptive deformable convolution kernel; The elliptical attention mechanism module generates an elliptical spatial constraint using iris anatomical priors for the matching of the attention mask and the biometric distribution; S3. Train and optimize the lightweight encoder-decoder network; Perform multi-task loss calculation, introduce an elliptical parameter regression constraint on the basis of the segmentation loss to design a loss function, and optimize the segmentation result; S4. Perform fine refinement of the elliptical parameters through the post-processing module, and perform elliptical fitting optimization on the segmentation result.

2. The iris pupil real-time segmentation method according to claim 1, wherein The preprocessing using dynamic occlusion enhancement is specifically to preprocess the original iris image through a combination strategy of random region erasing and elastic deformation.

3. The iris pupil real-time segmentation method according to claim 1, wherein The deformable convolution module adopts a method of adaptively deforming the convolution kernel according to the iris edge curvature. The mathematical expression of the deformable convolution kernel is specifically as follows: x l+1 = ReLU(BN(DefromConv(DefromConv(x l )))) Among them, p represents the spatial position coordinates of the output feature map; K is the sampling points of the 3×3 convolution kernel; Δp k ∈R 2 is the learnable offset; f offset is the offset prediction function; Through the dynamic adjustment of Δp k the convolution kernel can adapt to the curvature features of the iris texture.

4. The iris pupil real-time segmentation method according to claim 1, wherein, The elliptical attention mechanism module generates an elliptical heat map through elliptical parameter prediction, specifically as follows: First, parameterize the ellipse as follows: Θ = {c x ∈ [0, 1], c y ∈ [0, 1], a ∈ [0, W / 2], b ∈ [0, H / 2], θ ∈ [-π, π]} The parameter prediction process is as follows: Among them, F avy is the feature vector after global average pooling; Secondly, generate the elliptical heat map; The elliptical membership function is as follows: The rotation coordinate transformation is as follows: x′ θ =(x - c x W)cosθ+(y - c y H)sinθ y′ θ = -(x - c x W)sinθ+(y - c y H)cosθ Finally, perform feature enhancement, fusing the elliptical heat map and the original features: x attn = x ⊙ σ(Conv 3×3 (Concat(AvgPool(x), E(x, x)))) where ⊙ is element-wise multiplication.

5. The iris pupil real-time segmentation method according to claim 1, wherein The loss function adopts a dual-branch decoder combined with a joint loss function: The segmentation head outputs pixel-level probabilities: The elliptical parameter regression head predicts iris and pupil parameters: The joint loss function is: where λ1, λ2, λ3, and λ4 represent balance coefficients.

6. The iris pupil real-time segmentation method according to claim 5, characterized in that The joint loss function is specifically: Cross-entropy loss: Dice loss: Elliptical parameter regression loss: Shape consistency loss: Among them, represents the cross-entropy loss, represents the Dice loss, represents the elliptical parameter regression loss, represents the shape consistency loss.

7. The iris pupil real-time segmentation method according to claim 1, characterized in that The post-processing module adopts an elliptical constraint optimization function, specifically as follows: Among them, Ψ k represents iris pupil ellipse parameter extraction, is the Gaussian prior distribution, and ∑ is the credibility matrix of the prediction parameters.

8. An iris pupil real-time segmentation system applied to the method according to any one of claims 1-7, characterized in that, It includes: A lightweight encoder, including a deformable convolution module and an elliptical attention mechanism module, for feature extraction; A decoder for multi-scale feature fusion, including a convolutional layer, an elliptical parameter prediction head, and a segmentation head, for multi-scale feature fusion; An elliptical parameter-guided post-processing module for optimizing the iris segmentation prediction result.

Images