Technological developments in medical science grows rapidly; as a result, it becomes very complicated to identify the sex from facial image. So, to recognize the sex from facial image will become a vital problem and leads security issues. Consequently, this issue switches the attitude of signal and image processing researchers to exploit the problem; numerous theories and model have been developed to mitigate these issues. However, most approaches used complex image processing algorithms and make use neural networks to study the image performance and increases the delay thereby reduces the overall throughput.
Characteristics of the human face may be used to get information about a person's age, gender, emotional state, and ancestry
Real Convolution neural network (RCNN) has widespread success in various applications, however correlations between convolution kernels are rarely considered, and i.e., no specific link or connection is created among convolution kernels. Moreover, realvalue recurrent neural networks (Real RNNs) acquire the correlations by connecting and learning the weights of convolution kernels and the training difficulty is much increased.
Moumen et al ^{10} used Complex algebra and quaternion algebra to increase performance during model the connections between convolution kernels. Subsequently, lot of effort is required to design neural networks to work in the complex, quaternion, and octonion spaces, which are outside the existing domain. In
In this paper an enhanced algorithm called hybrid of HDSON and Bidirectional Associative Memory Model(HHBM). In this approach we make use of HDSON to consider the facial image and BAM to improve the storage capacity of the system.Consequently, a hybrid model is developed having both the properties of HDSON and BAM and improves the performance of the existing systems
The rest of the paper is organized as follows, in section 2 we are presenting our proposed model, section 3 presents results and discussion and finally we are concluding our paper in section 4.
The proposed model presented has the following folds.
1 e1 e5 e3 e4 e6 e7 e2 1 1 e1 e5 e3 e4 e6 e7 e2 e1 e1 −1 −e4 −e2 e5 −e7 e6 e3 e2 e2 −e3 e7 e1 −e7 −1 e2 −1 e3 e3 e2 −e6 −1 e7 e5 −e4 −e1 e4 e4 −e5 e1 −e3 e7 e1 e3 −e6 e5 e5 e4 −1 e6 −e1 −e3 e2 −e7 e6 e6 e7 e3 −e5 −e2 −1 −e1 e4 e7 e7 −e6 −e2 e4 −e3 e1 −1 e5
In the context of an octonion, it has been proven that this is identical to picking the atom with the highest correlation with the residual vector
Where,
Octonion and real convolution are denoted by the symbols
Where
Moreover, the octonion batch normalization layer's forward conduction can be calculated as:
Where
Since the 8 parts are zeromean then,
Here, the weight
Where:
And
The latter consists of m neurons and an output layer with s neurons. Weights
Where Re and
Furthermore, the realvalued delta rule is expanded upon as follows to help find the best values for network characteristics like weights and bias and can be given by
So by the applications of the above discussion the modified approach is therefore given by
There are now three stages in which the octonion output is fed. Multiple blocks with two convolution layers remain at each level. To ensure the expressiveness of the output features, the number of feature maps is steadily raised in all three stages. Next, an AveragePooling2D layer is used to speed up training, and then a fully connected layer named Dense is used to classify the data. There are three blocks in the deep octonion network model. The first block has a total of 10 residual blocks, the second has 9 convolution filters, and the final blocks have 32, 64, and 128 residual blocks. The batch size has been set at 64.


0.1 
(20, 60) 
0.01 
(0, 20) 
0.01 
(60, 80) 
0.001 
(80,110) 
0.0001 
(110, 120) 
Stochastic gradient descent and the crossentropy loss function are used in conjunction with deep octonion networks to train the final model. The Nesterov Momentum is set to 0.9 during descent to improve stability and speed convergence. Different learning rates are utilized in each network's epoch to improve network stability. To begin with, the learning rate is set at 0.01; it rises by a factor of 10 for the next 40 evaluations. Rather than training for 200 epochs, deep octonion networks are trained for only 120 epochs since their convergence speed is fast.
Kersaf is used on a PC running Ubuntu 16.04 with a CPU running at 3.40 GHz clock speed and 64 GB of RAM, and two NVIDIA GeForce GTX1080Ti graphics cards.
The performance of the classification algorithms was compared according to the criteria of accuracy, precision, and Fscore represented in equations (1720). In the equations, the male is defined as A, the male but predicted as female is considered as B, and the female but identified as male is denoted as C. Finally, the female is correctly identified as female and is represented as D.
Here the proposed model is validated by considering two different data ratios i.e. 70% Train 30% Test and 60% Train 40% Test. To validate our approach we consider various network models (i.e. LeNet, AlexNet, DenseNet and HDSON) as most of the researchers have used these networks to identify the gender. From the test analysis it has been observed that the accuracy of the proposed model achieved 98.84%, HDSON, LeNet, AlexNet and DenseNet achieved 96.51%, 90.71%, 93.69% and 95.71% respectively for 70%30%, and the for 60%40% the proposed model achieved 96.76 % whereas HDSON, LeNet, AlexNet and DenseNet achieved94.93 %, 89.57%, 91.49% and 93.76% respectively. The reason for the better performance of proposed model is that it used both the properties of HDSON and BAM. Therefore, the accuracy is improved for the recognition of gender using facial images. Similarly, in the analysis of S and p, the proposed model achieved 92% & 96% respectively, whereas the LeNet, AlexNet, DenseNet and HDSON achieved 86.67& 84.82%, 89.63% & 85.89%, 91.12% &89.5% and 91.01 % & 89.8 % respectively for different training and testing datasets. Furthermore, when the techniques were tested with FMeasure, the proposed model achieved 90% to 94% for 70:30. The same model achieved 89 % to 91% for 60:40. However, the HDSON achieved highest among all others i.e.91% for 70:30 and has 90.74% for 60:40.
This research proposes a new architecture to determine a person's sex by combining the features of HDSON and BAM. Here we employed octonion orthogonal matching pursuit technique to provide an alternate minimizationbased approach for the octonion sparse coding issue which is the expansion of the previous research and developed deep octonion networks (DONs). The OVNN model optimizes the octonion weight initialization by using ReLU as the activation function. As a result, the suggested algorithm is successful because it employed the AR data set, which contains photographs of faces captured by mobile phones from various angles and showed a wide range of emotions belonging to people of both sexes. According to the experimental analysis, the suggested approach provides rapid training with many pictures and obtained high classification accuracy.
The proposed approach improves the performance of gender detection 3%, as it combines the features of HDSON and BAM. The loss of extracted image is also being reduced, since proposed approach provides the qualitative properties of HADSON and BAM, hence acquired enhanced stored capability which sequentially minimizes the loss of image. Also, it integrates the qualities of HDSON and BAM to optimize time, storage, accuracy, sensitivity and precision of the extracted image. Moreover, the proposed algorithm is simple to operate, so provides the base to signal and image processing researchers to precede the work on this era. The proposed algorithm can be used in paramilitary forces to improve the security against opponents, even if they change their facial activity. Furthermore, the approach can be used to detect the missed persons and helps the local police in real time applications.