Iris Detection Neural Network Training Method

Abstract

The present invention discloses a method for training an iris detection neural network, which includes the following steps: extracting pre-stored training sample images and marked coordinates from a sample library; inputting the extracted training sample images and marked coordinates into a neural network model for training; calculating the network output based on the current network connection weights; calculating the inference loss Loss = 1 - IOU + r 2 / c 2 +(1 - a 2 / b 2 ) Calculate the inference loss Loss; a is the length of the short side of the inference box; b is the length of the long side of the inference box; c is the length of the diagonal of the inference box; r is the distance from the center point of the inference box to the center point of the marked box; determine whether the output loss Loss meets the preset conditions. If not, adjust the network connection weights according to the feedback algorithm, and re-infer and calculate the inference loss; if so, stop the training. This method can achieve a lower false recognition rate, a more accurate positioning effect, and a faster convergence speed when detecting the iris.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method for training an iris detection neural network. Background Art

[0002] The development of modern society has put forward higher requirements for the accuracy, security and usability of human identity recognition. At present, traditional photo recognition can no longer meet the requirements of the times. Biometric recognition has gradually replaced photo recognition and become the mainstream identity recognition method. Biometric recognition technology refers to using the inherent physiological or behavioral characteristics of the human body for identity recognition. Among them, iris recognition technology has random detailed features and texture features, and maintains a relatively high stability; it has the advantages of non-contact acquisition, etc., and has broad market prospects and scientific research value. Existing iris detection is realized based on deep learning technology, that is, through training a deep neural network with a large amount of data, so that the trained deep neural network has the ability to detect and accurately locate a certain type of object. The training of a deep neural network is that the neural network continuously learns a large amount of sample data, and then uses a preset error evaluation method to evaluate the learning effect of the neural network in each round, and adjusts the connection weights extracted by the network through a learning optimization algorithm for retraining until the error evaluation method believes that the training effect of the network has reached the expected requirements. The current mainstream method of calculating the training Loss by simply calculating the IOU between the inference box and the marked box cannot reflect the characteristics of the iris detection and positioning scenario. Therefore, how to develop a new method for training an iris detection neural network to overcome the above problems is the direction that those skilled in the art need to study. Summary of the Invention

[0003] The present invention provides a method for training an iris detection neural network, which can achieve a lower false recognition rate, a more accurate positioning effect and a faster convergence speed when detecting the iris.

[0004] A method for training an iris detection neural network includes the following steps:

[0005] Step 1: Extract the pre-stored training sample images and marked coordinates from the sample library;

[0006] Step 2: Input the training sample images and marked coordinates extracted in Step 1 into the neural network model for training;

[0007] Step 3: Calculate the network output based on the current network connection weights;

[0008] Step 4: Based on the loss function Loss = 1 - IOU + r 2 / c 2 +(1 - a 2 / b2 ) Calculate the inference loss Loss; where a is the length of the short side of the inference box; b is the length of the long side of the inference box; c is the length of the diagonal of the inference box; r is the distance from the center point of the inference box to the center point of the marked box;

[0009] Step 5: Determine whether the output loss Loss meets the preset conditions. If not, adjust the network connection weights according to the feedback algorithm and jump to Step 3 to re-infer and calculate the inference loss; if it meets, the training stops.

[0010] Preferably, in the above iris detection neural network training method: in Step 3, based on the function Calculate the output of each layer in the fully connected layer.

[0011] More preferably, in the above iris detection neural network training method: the preset condition in Step 5 is that the output loss Loss is less than a preset first threshold.

[0012] By adopting the above technical solution: the Loss between the inference box and the marked box is modified from simply calculating the IOU between the two to: not only calculating the IOU, but also calculating the center deviation, the length of the diagonal, and the aspect ratio of the inference box. Thus, the training result of the deep neural network can better meet the requirements of the iris detection and positioning task. Compared with the prior art, this method can have a faster training convergence speed, a higher correct recognition rate, a lower false recognition rate, and a more accurate and complete positioning effect in the application field of iris recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the working process of the present invention;

[0014] Figure 2 It is a schematic diagram of the first group of iris recognition comparison models in Embodiment 1;

[0015] Figure 3 It is a schematic diagram of the second group of iris recognition comparison models in Embodiment 1;

[0016] Figure 4 It is a schematic diagram of the third group of iris recognition comparison models in Embodiment 1. DETAILED DESCRIPTION OF THE INVENTION

[0017] As Figure 1 shown:

[0018] Embodiment 1: An iris detection neural network training method, which includes the following steps:

[0019] Step 1: Extract the pre-stored training sample images and marked coordinates from the sample library;

[0020] Step 2: Input the training sample images and marked coordinates extracted in Step 1 into the neural network model for training;

[0021] Step 3: Based on the function Calculate the output of each layer of the fully connected layer;

[0022] Step 4: Based on the loss function Loss = 1 - IOU + r 2 / c 2 +(1 - a 2 / b 2 ), calculate the inference loss Loss; where a is the length of the shorter side of the inference box; b is the length of the longer side of the inference box; c is the length of the diagonal of the inference box; r is the distance from the center point of the inference box to the center point of the marked box;

[0023] Step 5: Determine whether the output loss Loss meets the preset condition, that is, whether the output loss Loss is less than the preset first threshold. If Loss is greater than or equal to the preset first threshold, adjust the network connection weights according to the feedback algorithm and jump to Step 3 to re-infer and calculate the inference loss; if Loss is less than the preset first threshold, it means the training is successful, and the training ends at this time.

[0024] In the above steps:

[0025] The part of calculating the training Loss by calculating the IOU between the inference box (i.e., Predict box) and the marked box (i.e., Ground truth) in the prior art is retained, and several calculation items are further added:

[0026] As Figure 2 shown, it is a comparison schematic diagram of two models with the same IOU and a certain area of the inference box and the marked box, but different distances between the center points of the inference box and the marked box. Figure 2 The recognition scene model on the left in

[0027] is better than the recognition scene model on the right. Among them, r represents the distance between the two center points of the inference box and the marked box. The farther the distance between the two center points is, the larger r is, and the larger Loss is. Since in the training process of the deep neural network, the network always adjusts in the direction of reducing Loss, so as to ensure that the center points of the inference box and the marked box can get closer and closer during the training process. c represents the length of the diagonal of the inference box. The longer the length of the diagonal of the inference box is, the larger the area of the inference box is.

[0028] As Figure 3 shown, it is a comparison schematic diagram of two models with the same IOU and a certain area of the inference box, but different area ratios of the inference box and the marked box. Figure 3 The recognition scene model on the left in

[0029] It can be seen that on the premise of the same IOU, it is better for the inference box to be larger rather than smaller. Therefore, the present invention helps the deep neural network to select the result with a relatively larger inference box area during the training process on the premise of satisfying that the larger the IOU, the better. Since a larger area of the inference box indicates information redundancy within the framed range, while a smaller area of the inference box indicates information loss within the framed range. Comparing the two, it is better to have information redundancy rather than information loss, because redundant information can be filtered out through subsequent feature extraction processes, but once information is lost, it cannot be compensated for.

[0030] Such as Figure 4 shown in the comparison schematic diagram of two models with different length ratios between the short side and the long side of the inference box on the premise of the same IOU and a certain area of the inference box. Figure 4 The recognition scene model on the left in it is superior to the recognition scene model on the right.

[0031] Among them, a represents the short side of the inference box, b represents the long side of the inference box, and the closer the lengths of the short side and the long side of the inference box are to each other, the smaller the Loss. It helps the deep neural network to select the result with a shape closer to a square for the inference box during the training process on the premise of satisfying that the larger the IOU, the better. Since the iris itself can be considered to be in a circular shape, the closer the shape of the inference box itself is to a square, the closer it is to the shape of the iris itself.

[0032] Through batch testing with big data, after adopting the iris detection method provided by the present invention, the average degree of coincidence between the inference box and the marked box reaches more than 90%, while the average coincidence rate between the iris inference box and the marked box obtained by using the existing object detection method is generally only 75%-80%.

[0033] Embodiment 2:

[0034] An iris detection neural network training method, which includes the following steps:

[0035] Step 1: Extract the pre-stored training sample images and marked coordinates from the sample library;

[0036] Step 2: Input the training sample images and marked coordinates extracted in Step 1 into the neural network model for training;

[0037] Step 3: Calculate the network output based on the current network connection weights;

[0038] Step 4: Based on the loss function Loss = 1 - IOU + r 2 / c 2 +(1 - a 2 / b 2) Calculate the inference loss Loss; where a is the length of the short side of the inference box; b is the length of the long side of the inference box; c is the length of the diagonal of the inference box; r is the distance from the center point of the inference box to the center point of the marked box;

[0039] Step 5: Determine whether the output loss Loss meets the preset conditions. If not, adjust the network connection weights according to the feedback algorithm and jump to Step 3 to re-infer and calculate the inference loss. When the number of times of jumping to Step 3 to re-infer and calculate the inference loss reaches the preset second threshold, the training ends automatically.

[0040] In the case of Embodiment 2, it shows that although the training reaches the preset number of times, the loss still exceeds the preset threshold. At this time, it indicates that the network has underfitting. At this time, stop the training and adjust the training samples to retrain.

[0041] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for training an iris detection neural network, characterized in that, It includes the following steps: Step 1: Extract the pre-stored training sample images and marked coordinates from the sample library; Step 2: Input the training sample images and marked coordinates extracted in Step 1 into the neural network model for training; Step 3: Calculate the network output based on the current network connection weights; Step 4: Based on the loss function , calculate the inference loss Loss; The a is the length of the short side of the inference box; the b is the length of the long side of the inference box; the is the diagonal length of the inference box; the r is the distance from the center point of the inference box to the center point of the marking box; Step 5: Determine whether the output loss Loss meets the preset conditions. If not, adjust the network connection weights according to the feedback algorithm and jump to Step 3 to re-infer and calculate the inference loss; if it meets, the training stops; When the number of times of jumping to Step 3 to re-infer and calculate the inference loss reaches the preset second threshold, the training ends automatically.

2. The iris detection neural network training method according to claim 1, characterized in that: In step 3, based on the function calculate the output of each layer of the fully connected layer.

3. The iris detection neural network training method according to claim 1, characterized in that: The preset condition described in Step 5 is that the output loss Loss is less than the preset first threshold.

Images