CN-121982752-A - Depth separable convolution-based lightweight fingerprint living body detection method, storage medium and device

CN121982752ACN 121982752 ACN121982752 ACN 121982752ACN-121982752-A

Abstract

The invention discloses a lightweight fingerprint living body detection method, storage medium and device based on depth separable convolution, and belongs to the field of image recognition and classification. The method comprises the steps of preprocessing an acquired original fingerprint image, constructing a lightweight backbone network which is formed by connecting five feature extraction layers in series to extract features, introducing differential residual connection in a second layer to a fifth layer to strengthen the features and relieve gradient attenuation, extracting context features in parallel through a multi-scale cavity convolution module and fusing, and finally finishing the discrimination of the living body through global average pooling and a classifier. The invention replaces standard convolution by depth separable convolution in the deep layer of the network and combines residual connection and multi-scale feature fusion mechanism, thereby obviously reducing the parameter number and calculation complexity of the model while ensuring high detection precision, realizing high-efficiency light weight of the model, and being particularly suitable for real-time and safe fingerprint living body detection on an embedded or mobile terminal with limited calculation resources.

Inventors

GONG ZHIGANG
QIU YILIN
ZHU ZEQING
GONG ZHIWEI
PENG DI

Assignees

重庆大学
圆月辰界(重庆)人工智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260209

Claims (8)

1. The lightweight fingerprint living body detection method based on depth separable convolution is characterized by comprising the following steps of: S1, preprocessing an original fingerprint image acquired by an electronic fingerprint sensor to obtain a standardized image; S2, constructing a sequence of feature images extracted from each layer of a lightweight backbone network extraction standardized image, wherein the sequence of feature images is formed by sequentially connecting five feature extraction layers in series; s3, introducing differentiated residual connection into each of the second to fifth feature extraction layers of the lightweight backbone network, enhancing the features, and finally outputting an enhanced feature map through the fifth feature extraction layer; S4, extracting multi-scale context features of the enhanced feature map by using convolution kernels with different scales and hollowness in parallel, and performing channel splicing to obtain multi-scale features; s5, inputting the multi-scale features into a global average pooling layer to be laminated into feature vectors, then connecting the full-connection layer with a Softmax classifier to judge whether the fingerprint is living or not, and outputting a detection result.
2. The depth separable convolution based lightweight fingerprint living body detection method according to claim 1, wherein said preprocessing of step S1 comprises size normalization, pixel value normalization, noise suppression and smoothing.
3. The depth separable convolution-based lightweight fingerprint living body detection method as recited in claim 1, wherein the lightweight backbone network is composed of five feature extraction layers sequentially connected in series, wherein: the first feature extraction layer is formed by series batch normalization (BatchNorm) of standard convolution blocks and series ReLU activation functions; The second feature extraction layer is formed by stacking maximum pooling series standard convolution; The third feature extraction layer is a standard convolution stack; the fourth feature extraction layer is a standard convolution stack; The fifth feature extraction layer is a standard convolution stack.
4. A depth separable convolution based lightweight fingerprint living body detection method according to claim 3, wherein said fourth and fifth feature extraction layers are depth separable convolution stacks.
5. A depth separable convolution based lightweight fingerprint living body detection method according to claim 3, wherein the number of stacked layers of said second to fourth feature extraction layers increases in sequence.
6. The lightweight fingerprint living body detection method based on depth separable convolution according to claim 1, wherein the step S3 specifically comprises: s301, for a first feature extraction layer, direct output of a feature map without introducing residual connection; s302, introducing direct residual connection and expansion convolution residual connection in parallel for the second to fourth feature extraction layers, and superposing and fusing the input features, the standard convolution features and the expansion convolution features; and S303, outputting the enhanced feature map by a direct residual connection mode for the fifth feature extraction layer.
7. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the depth separable convolution-based lightweight fingerprint in vivo detection method of any one of claims 1 to 6.
8. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the depth separable convolution based lightweight fingerprint biopsy method of any one of claims 1 to 6.

Description

Depth separable convolution-based lightweight fingerprint living body detection method, storage medium and device Technical Field The invention relates to a lightweight fingerprint living body detection method, a storage medium and a device based on depth separable convolution, belongs to the field of image recognition and classification, and is particularly suitable for lightweight fingerprint living body detection based on depth separable convolution. Background With the rapid development of information technology, biometric identification is becoming a main identity authentication mode in the fields of financial payment, intelligent access control, enterprise sentry verification, public security and the like. Fingerprints are considered one of the most mature, most reliable biometric features due to their uniqueness, stability and lifetime invariance. However, the widespread use of fingerprint recognition systems has also created new security challenges, and attackers often make counterfeit fingerprint films from materials such as silica gel, latex, gelatin, etc. to fool the recognition system, threatening property security, personal safety, or even judicial fairness. To combat counterfeit fingerprint attacks, fingerprint in vivo detection (Presentation Attack Detection, PAD) techniques are becoming a research hotspot. From early hardware sensor detection and manual feature extraction based on expert knowledge, to a traditional machine learning method based on texture and optical features, to a deep learning end-to-end model of the current mainstream, the detection performance is continuously improved. Especially, the deep learning model can automatically learn the slight difference between the real fingerprint and the fake fingerprint, and the detection precision is remarkably improved. However, the existing deep learning model has the problems of large parameter quantity, high calculation complexity, low reasoning speed and the like, and is difficult to be deployed on embedded equipment, mobile terminals or judicial evidence obtaining terminals with limited resources in real time. Therefore, how to realize the light weight and real-time reasoning of the model while ensuring the detection precision has become a key bottleneck for the floor application of the fingerprint living detection technology. Disclosure of Invention In order to solve the problems of large model volume, high inference delay, strong dependence on computational power resources and difficult stable operation on an electronic fingerprint acquisition terminal and embedded equipment in the prior fingerprint living detection method in practical application, the invention provides a lightweight fingerprint living detection method, a storage medium and equipment based on depth separable convolution, aiming at fingerprint images acquired by an electronic fingerprint sensor, by constructing a depth feature extraction network with a residual error connection structure and multi-scale sensing capability, the model parameter scale and the calculation complexity are obviously reduced while the living body detection precision is ensured, so that the comprehensive requirements of the fingerprint living body detection on instantaneity, stability and resource occupation in an actual application scene are met. In order to achieve the above purpose, the present invention provides the following technical solutions: the lightweight fingerprint living body detection method based on depth separable convolution is characterized by comprising the following steps of: S1, acquiring an original fingerprint image of an electronic fingerprint sensor Preprocessing to obtain standardized image; S2, constructing a lightweight backbone network extraction standardized image formed by sequentially connecting five feature extraction layers in seriesIs a sequence of extracted feature maps of each layer of (a)、、、、; S3, introducing differentiated residual connection into each of the second to fifth feature extraction layers of the lightweight backbone network, enhancing the features, and finally outputting an enhanced feature map through the fifth feature extraction layer; S4, extracting the enhanced feature map by using convolution kernels with different scales of void ratios in parallelPerforming channel splicing (connection) on the multi-scale context characteristics to obtain multi-scale characteristics; S5, multi-scale featureAnd (3) inputting a global average pooling layer (GAP) to compress as a feature vector, then connecting a full connection layer and a Softmax classifier to judge whether the fingerprint is living or not, and outputting a detection result. Further, the preprocessing operation in step S1 specifically includes performing size normalization processing on the original fingerprint image to uniformly map the fingerprint images acquired under different resolution conditions to a fixed input size required by the network, performing pixel value normalizati