Detection of Breast Cancer in Mammogram Images Using Multi Attention Feature Extraction with Hybrid RSA Based AlexNet

Large connected gadgets link to the Internet of Things (IoT), an interactive paradigm that is constantly increasing and enabling communication between people and things in our current digital environment. IoT includes all medical devices that gather patient health information, including smartphones, smartwatches, wearable wristbands, integrated surgical instruments, and other flexible health monitoring gadgets [1]. This technology is used in both homes and hospitals to remotely monitor patients, facilitating the early diagnosis and treatment of serious illnesses and medical problems. Remote patient monitoring in hospital settings has the potential to drastically lower needless doctor visits, hospital stays, readmissions, and total healthcare expenses [2]. Despite the difficulties and barriers encountered in the past, IoT technologies have revolutionized the world of medical applications [3]. These IoT devices have the potential to provide large amounts of biomedical data, which may be very helpful in the creation of systems that automatically gather medical data [4]. As IoT devices are connected, big data, together with cutting-edge Machine Learning (ML) technology, plays a crucial role in improving current healthcare systems' diagnosis, treatment, and decision-making [5].

The Internet of Everything (IoE), which incorporates symptomatic treatments and monitoring along with tracking of patients, has attracted research attention as a result of IoT in biomedical applications [5]. The World Health Organization (WHO) ranks breast tumors as the most common worldwide reason for death in women. Early identification of breast tumors in women is the greatest approach to saving lives and reducing healthcare expenditures [6]. In general, cancers are classified as either benign or malignant. Benign tumors don't pose a life-threatening hazard, but they may make women more likely to develop breast cancer. On the other hand, malignant tumors are cancerous and need to be treated right away. According to research on breast cancer screening, 20% of women had malignant tumors [7]. Breast tumors may be identified over an extended duration of time by utilizing a number of methods, including X-rays, ultrasound, mammography, magnetic resonance imaging (MRI), positron emission tomography (PET), and computed tomography [8]. Breast cancer in its early stages may be assessed by screening, which finds the disease when small breast symptoms appear. Mammography, clinical breast exams, and other methods are additional methods for breast screening. In mammography, doctors utilize brief-intensity X-rays to scan the breasts for anomalies. There are two other imaging techniques to check for breast cancer issues: ultrasound and MRI scans [9]. A "one-size-fits-all" screening strategy runs the risk of underdiagnosing breast tumors, particularly in women with dense breasts. Breast density, which is the ratio of fibroglandular tissue to fatty tissue in the breast, affects the sensitivity of mammography [10]. Mammography has established as the most accurate approach for identifying breast tumors because of its capacity and affordability to meet medical needs. The major technique used by doctors to make diagnoses is mammography analysis, although this approach is subject to bias and physician fatigue [11].

The mammography technique has a poor rate in detecting breast tumors, which is unfortunate. Depending on the type of tumors, the general size of the female breasts, and the elder age of the victim, it may give false-negative results of 5% to 30% [12]. As a result, mammography uses low-dose radiography, which enables viewing of the internal breast tissue. Breast cancer is detected using a variety of signal processing techniques, such as curvelet transform, microwave imaging, and ultrasound imaging. The traditional approach for categorizing medical infections, such as skin blemishes, breast lumps, and brain tumors, is based on pattern recognition [13]. The mammography characteristics for breast cancer are manually retrieved, and they are then sent inside an ML classifier for categorization. Obtaining a proper classification, however, remains challenging because of many imaging difficulties and alterations in tumor regions. As a result, deep learning (DL) has played a major part in the past ten years in the detection and classification of medical infections, particularly for breast cancer [14].

Current techniques for identifying breast cancer from mammography pictures mostly depend on deep learning architectures or classic machine learning algorithms. Although these techniques have showed promise, they frequently fail to capture the complex patterns and nuanced characteristics typical of breast cancer in its early stages. Additionally, they can be less effective in clinical situations due to noise and fluctuations in image quality.

By utilizing cutting-edge approaches in deep learning and attention mechanisms, the suggested method, "Detection of Breast Cancer in Mammogram Images using Multi-Attention Feature Extraction with Hybrid RSA-based AlexNet," seeks to overcome these drawbacks. The major contributions in this paper are:

• Preprocessing the mammography images to eliminate noise is a vital initial step in our methodology.
• In order to extract pertinent characteristics from the images, this paper uses the MAFNet. It uses the SMO approach to optimize the learning rate in MAFNet, resulting in effective feature extraction.
• For classification, the AlexNet model is used. The ACO-RSA hybrid optimization strategy is utilized to tune the AlexNet classification technique’s hyperparameters.
• Performance measures such as accuracy (ACC), recall (RC), precision (PR), and F1 are used to assess overall results.

Organization of the work

Section 2 reviews the relevant works, Section 3 offers a brief description of the proposed model, Section 4 illustrates the results and testing analysis, and then Section 5 delivers the results in pictorial form and tabulation with conclusion.

Figure 1 depicts the stages involved in putting the proposed approach into action. The section (workflow) contains image preprocessing, feature extraction using SMO based MAFNet and hybrid optimization based AlexNet feature classification.

tu_pian_1.png

Figure 1. Workflow

3.1 Dataset description

The study conducts an in-depth analysis utilizing data extracted from 1459 mammograms sourced from the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM). The CBIS-DDSM dataset represents an enhanced version of the Digital Database for Screening Mammography (DDSM), which is widely recognized and utilized in breast cancer research and diagnostics. This upgraded manifestation, the CBIS-DDSM dataset, comprises a diverse range of mammographic images annotated with detailed information regarding various abnormalities and lesions.

Within the CBIS-DDSM database, a comprehensive array of mammogram images is categorized based on the presence and nature of abnormalities. Specifically, the dataset encompasses 398 images depicting benign calcifications, 417 images illustrating benign masses, 300 images showcasing malignant calcifications, and the remaining images displaying malignant masses. These classifications provide a nuanced representation of different pathological conditions and abnormalities commonly encountered in mammography screenings.

All images contained within the CBIS-DDSM dataset adhere to a standardized format, with dimensions set at 224 × 224 pixels and encoded in the RGB (Red, Green, Blue) color space. This uniformity ensures consistency in image resolution and format across the entire dataset, facilitating streamlined data preprocessing and analysis procedures. Table 1 in the study provides a comprehensive overview of the dataset, summarizing key characteristics and distribution statistics pertaining to the different categories of abnormalities represented within the CBIS-DDSM database. This summary serves as a valuable reference for researchers and practitioners involved in the field of breast cancer detection and diagnosis, offering insights into the composition and diversity of the dataset under scrutiny.

Overall, the utilization of the CBIS-DDSM dataset in the study underscores the importance of leveraging high-quality, annotated data repositories to drive advancements in medical imaging research, particularly in the context of breast cancer detection and diagnosis. The dataset's richness in pathological variations and standardized image attributes empowers researchers to develop and validate robust algorithms and methodologies aimed at improving the accuracy and efficiency of breast cancer screening and prognosis.

Table 1. Dataset description

After receiving the dataset, several photos that were just artifacts labeled as malignant masses or blank images were discovered. ACC might suffer as a result of these photographs. 17 photos were therefore manually eliminated. 327 malignant mass mammograms remained in the photos after the photographs had been removed, whereas the number of mammogram images in the other categories remained steady. As a consequence, 1442 photos in a dataset with four distinct classifications are produced. Examples of four classes are shown in Figure 2, along with information on their traits and artifacts.

tu_pian_2.png

Figure 2. Dataset (CBIS-DDSM) comprising mammograms having four classifications in which multiple artefacts appear in each class

Figures 2(a) and (b) both include a little label at the bottom edge of the picture, a huge label in Figure 2(b), a pair of straight lines linked towards the image's breast region in Figure 2(c), and a straight line at the bottom edge of Figure 2(d).

3.2 Preprocessing

Before inserting the pictures inside a neural network for analysis, image pre-processing thought to be the most crucial operation to attain a reasonable level of ACC and shorten the computing time of a model. It is complicated to an artificial neural network technique in order to categorize mammograms without first using pre-processing methods. As a result, picture pre-processing is done first. The studies [22, 23] outlined many pre-processing methods that may be used without degrading the quality of the original picture. Two further research [24] improved the contrast of mammography using tried-and-true techniques.

In this part, it is explained how mammography quality and quantity might be improved. This covers picture improvement, background removal, and artifact removal. The primary processes, related sub-processes of picture pre-processing phase are shown in Figure 3. Artefact removal, remove line, picture enhancement, and testing are the primary procedures.

tu_pian_3.png

Figure 3. Phase chart of image pre-processing where every process is depicted with a structure of blocks

Block (A): A variety of approaches are used to remove any anomalies seen in original mammograms as: binary masking of the image, morphological opening (second process), and largest contour identification that produce artefact at ease mammograms; In block (B) lines are connected with the female breast portion of the mammograms eliminated via different techniques: in Range Operation(the early process) following Gabor filter, Morphological opening, which is the third process and then Invert mask manipulation; (C) image enhancement through further Gamma correction following, 1^st then 2^nd contrast limited adaptive histogram equalization and Green Fire Software with a Blue Filter, which is ImageJ; subsequently in the block (D) the ACC of modified pictures has been examined by image statistical evaluation.

To get a more accurate result, artifacts from the mammograms are first eliminated, then certain techniques (binary masking of images, followed by process of morphological opening process, and biggest contour identification, the final phase) are used. Second, the vertical line connected to the breast region is removed using the "remove line" step. In this study, techniques including Range operation, Gabor filter morphological operation subsequently inverse masking methods were applied. The subprocesses of opening and dilatation are both part of the morphological operation. Thirdly, image enhancement is used to boost the contrast and luminosity of the initial pictures in order to improve the visibility of the malignant condition. Gamma correction, which is depicted in the block image [25], CLAHE (1st process) [26], CLAHE (2nd process), and the green-coloured fire blue ImageJ filtering [27] are the subprocesses of this stage2. An increase in visibility may be seen after using CLAHE. Then, CLAHE is used once again to boost the contrast. In other words, the method involves applying CLAHE twice. CLAHE should not be used a third time since doing so might distort the contrast level and impair the ability to see fine details in mammograms. Finally, in the verification step, evaluation methods are performed to the processed pictures to analyze the outcomes, including PSNR (peak signal-to-noise ratio), MSE (mean square error), RMSE (root mean squared error), SSIM (structural similarity index) and histogram analysis. Figure 3 depicts the entire pre-processing pipeline, with the result of each stage serving as the input for the next stage.

3.3 Feature extraction using SMO based MAFNet

3.3.1 MAFNet architecture

As a feature extractor, multi-attention fusion CNN (MAFNet) technique is used in this research. Four convolution models are present in the centre of MAFNet's $1 \times 1 \times 1$ convolution layers [28], while three convolution frameworks are found throughout all the modules. And the distribution of the convolution module follows a symmetrical pattern similar to [29]. Finally, there are 26 convolution modules in the FC layer [30, 31]. Following that, the pictures are supplied as input, and the first $1 \times 1$ convolutional method is carried out. 4 convolution algorithms are subsequently passed for four times again. The contextual transformer (CoT) block of the primary ResNet (Residual Network) convolution structures replaces $3 \times 3$ convolutional operation. The characteristics recovered are next provided as feed to FC levels (layers), which carry out the function of "classification" in neural network approach, following the early convolution, pooling, and excitation procedures [32]. Figure 4 depicts the MAFNet model's architecture. To assess the likelihood of categorization, it also uses the Softmax function. The outcome resulting from the classifier's results is finally provided as below.

$Softmax\left(z_j\right)=\frac{e^{z_i}}{\sum_{i=1}^n e^{z_i}}$             (1)

tu_pian_4.png

Figure 4. Architecture of MAFNet

Eq. (1) exhibits the total number of groups or classes, $z_j$ indicates result value for j^th node, $z_i$ denotes the result value of i^th node, while e indicates the fundamental constant, which may be computed as follows.

$S L_{C E}=-\sum_{i=1}^N l_i \log p_i$           (2)

The variable li in Eq. (2) refers to the thermal encoding of individual tags i from the set $\mathrm{i}(0, \ldots, \mathrm{N}-1)$. When li equals 1, it implies that the ith tag is present among the labels, while all other labels are set to zero. N also reflects the total number of tags present. The prediction probability associated with the i^th label is represented by p_i which corresponds to the Softmax score when the target label is i.

This needs to be observed that a lower learning rate could conclude in less rapid integration, while the maximum learning rate might end up resulting in a continuous fluctuation of loss function. When trying to construct a system with outstanding ACC and optimal variables accompanied with quick learning, an adjustable learning rate technique is used, which allows the learning rate to be adjusted every thirty epochs, where calculation appears below [33].

$l r=l r_0 \times 0.1^{\frac{epoch}{30}}$             (3)

In Eq. (3), lr signifies the current learning rate, lr₀ indicates the main learning rate, while epoch denotes the whole number of training cycles.

3.3.2 Learning rate optimization using SMO

Spider Monkey Optimization (SMO) plays an essential role in optimizing the learning rate (lr), as specified in Eq. (3). The SMO technique is a metaheuristic approach that was inspired by the social behavior of spider monkeys. This novel approach to problem solving draws on the collective behaviors of spider monkeys, integrating features of fission and swarm intelligence, as mentioned in the studies [34, 35]. Spider monkeys reside in communities between 40 and 50. A leader divides the responsibility of finding food in a community. Typically, a spider monkey swarm consists of 40 to 50 individuals, and the leader divides out the responsibility of finding food within an area. The worldwide leader of the swarm in the case of a food scarcity is always a leading female, which results in changeable smaller groupings. Based on the food supply in a particular area, the group size is determined. The spider monkey's size is closely correlated with the amount of food available. The required requirements are satisfied by the SMO-based approach of swarm intelligence (SI).

Smaller groups are formed to share the spider monkeys' foraging tasks. Self-organization: The group size is used to determine the food availability requirement. Intelligent foraging activity results in an intelligent choice. Swarm is used to start the food hunt. The distance between persons serving as food sources is computed. The distances between the members of the food groups change the locations for choosing. It is calculated how far apart individual is from their food supply.

tu_pian_5.png

Figure 5. Flowchart of SMO

For its six iterative collaborative stages, the SMO approach relies on trial and error: global leader decision phase, learning phase, local leader learning phase, local leader decision phase, global leader and local leader phase. ACO algorithms, inspired by the foraging behavior of ants, excel in exploring complex search spaces and identifying optimal solutions through iterative interactions. The collaborative framework of hybrid ACO-RSA enhances the robustness and generalizability of AlexNet by mitigating the risk of model bias and overfitting. By leveraging diverse optimization strategies, the model can effectively adapt to variations in mammogram images and generalize its diagnostic capabilities across diverse patient populations and imaging conditions.

In summary, the selection of SMO for MAFNet and the integration of a hybrid ACO-RSA approach with AlexNet reflect a strategic fusion of bio-inspired optimization techniques tailored to the unique requirements of breast cancer detection in mammogram images. These algorithmic choices contribute to the development of robust, efficient, and clinically relevant detection frameworks, ultimately advancing the field of medical image analysis and improving patient outcomes in breast cancer diagnosis and treatment. This approach's workflow is shown in Figure 5.

Below is a detailed explanation of the SMO approach.

Initializing

Population $P$ of spider monkeys are distributed using the SMO technique, where $S M_p$ stands for the $p-t h$ monkey in the population and $p=1,2 \ldots P$. The entire number of variables in a monkey is $M$, making them an M-dimensional vector. Using Eq (4), one potential response to each $S M_p$.

$S M_{p q}=S M_{minq }+U R(0,1) \times\left(S M_{maxq }-S M_{minq }\right)$              (4)

where, $S M_{p q}$ is $p t h S M$ of $q t h$ dimension. $S M_{p q}$ lower and upper bounds are $S M_{\text {ming }}$ and $S M_{\text {maxg }}$ in the $q t h$ direction for random number of $U R(0,1)$ uniform distribution is within the range of $[0,1]$.

Local Leader Phase (LLP)

LLP is an important part of the process in which SM, the local leader, changes the current location based on previous occurrences involving local group members. This change is done to improve the applicability of the new site in comparison to the prior one. The decision to relocate SM is conditional on the new site's fitness value exceeding that of the existing location. Eq. (5) provides the formula for updating the location of the $p t h$ SM within the lth local group.

$\begin{gathered} SMnew_{p q}=S M_{p q}+U R(0,1) \times\left(L L_{l q}-S M_{p q}\right)+ U R(-1,1) \times\left(S M_{r q}-S M_{p q}\right)\end{gathered}$             (5)

The symbol for the $q$ th dimension's $l$ th local group leader location is $L L_{l q}$.The qth dimension's $l$ th local group of the lth $S M$ is chosen at random and is designated as $S M_{r q}$, where $r, p$.

Global Leader Phase (GLP)

The local group members' input and the global leader's insights are used to guide the location improvement process, which is carried out in accordance with the LLP. Eq. (6) is used to calculate the new location data.

$\begin{gathered}SMnew_{p q}=S M_{p q}+U R(0,1) \times\left(G L_{l q}-S M_{p q}\right)+U R(-1,1) \times\left(S M_{r q}-S M_{p q}\right)\end{gathered}$              (6)

where, $q$ th dimension of global leader location is denoted as $G L_{l q}$ and an arbitrarily selected index is $q=1,2,3, \ldots M$.

In this step, the SM fitness determines probability prbp. Based on the probability value, the location of $S M_p$ is updated, and a better site’s candidate can allow to a variety of opportunities to enhance convergence. Eq. (7) provides the probability calculation results.

$p r b_q=\frac{f n_p}{\sum_{p=1}^N f n_p}$             (7)

where, $p t h$ $S M$, the symbol for fitness value is $f n_p$. A comparison is made between the SMs' new location fitness and their prior location. The location's top fitness value is taken into account.

Global Leader Learning (GLL) Phase

The global leader location has been updated by using a greedy approach. New spider monkey’s position is based upon world leader position for optimum population viability. The ideal site is used to apply the global leader. There is a 1 increment added to Global Limit Count in the event of updates.

Local Leader Learning (LLL) Phase

The local group and a greedy selection method are used to update the position of the local leader. The SM location is modified to guarantee the local leader's ideal placement inside a particular local group. The local leader's position is then precisely adjusted to ensure success. The local limit count is raised by one if no new entries are found.

Local Leader Decision (LLD) Phase

Local group candidates alter location at random in accordance with step 1 if the local leader can’t able to alter the position or uses the existing data from the global and local leaders based on the Eq. (8).

$\begin{array}{r}SMnew _{p q}=S M_{p q}+U R(0,1) \times\left(G L_{l q}-S M_{p q}\right) +U R(0,1) \times\left(S M_{r q}-L L_{p q}\right)\end{array}$             (8)

Global Leader Decision (GLD) Phase

If the position is not changed for a global leader up to the GLL, according to the preferences of the global leader, the populace is split into smaller sections. After receiving a maximum number of groups (MG), the splitting procedure begins. Each time, a local authority figure is chosen for recently established group. The greatest number of permissible sets is formed, and the global leader remains in their position until the pre-fixed permitted limit has been attained. At that point, the global leader seeks to combine all permitted groups into an individual group. The largest number of permissible chains is formed, and the global leader holds on to that position until the pre-fixed maximum has been reached. At that point, the global leader seeks to combine all permitted chains to unite. The following are the SMO evaluating control specifications:

Rate of perturbation (pr);

Maximum number of groups (MG);

Global Leader Limit;

Rate of Local Leader Limit.

Gaussian Mutation

In difficult iterative optimization situations, the SMO approach gets stymied in a local optimum. The algorithm solution value does not vary throughout iteration. This strategy leaves the location of the local optimum and includes stochastic perturbation and the Gaussian mutation before continuing on executing the creed in order to enhance the algorithm probability and algorithm deficiency. Eq. (9) displays the Gaussian mutation.

$x^{i, i t e r+1}=\left\{\begin{array}{cc}x^{i, i t e r}+rand & \text { if } r \geq 0.2 \\ x^{i, iter } \times Gaussian(\mu, \sigma) & \text { otherwise }\end{array}\right\}$            (9)

where, $r j$ is random fluctuation and $r a n d$, a randomized number between the interval of $[0,1]$. The dispersion of Gaussian variance is given in Eq (10).

$Gaussian(\mu, \sigma)=\left(\frac{1}{\sqrt{2 \pi \sigma}}\right) \exp -\left(\frac{(x-\mu)^2}{2 \sigma^2}\right)$        (10)

where, $\sigma^2$ and $\mu$ are denoted for variance and mean value.

3.4 Classification using hybrid optimization based AlexNet

3.4.1 AlexNet model

When features that had been extracted, the AlexNet CNN DL architecture has been used in the proposed study to classify breast cancer. The network's architecture has 5 convolution layers, which are followed by an average of 3 pooling layers, making it deeper than ordinary CNN. To prevent data overfitting, a dropout percentage of 0.5% has been assigned to the entirely linked layers present. The following elements make up the architecture:

• 1 Convolution with 11 × 11 kernel size (1CONV)

• Rectified Linear Unit Layer Activation (RELU)

• 3 Maximum Pooling (3 × 3)

• 1 Maximum Pooling (4 × 4 kernel)

• Rectified Linear Unit Layer Activation (RELU)

• Rectified Linear Unit Layer (RELU)

• Response Normalization Layer

• Rectified Linear Unit Layer (RELU)

• Rectified Linear Unit Layer Activation (RELU)

• 4 Convolution with 3 × 3 kernel size

• Rectified Linear Unit Layer Activation (RELU)

• 2 Convolution with 5 × 5 kernel size (2CONV)

• Fully Connected Layer (4096 nodes)

• 2 Maximum Pooling (3 × 3)

• Fully Connected Layer (4096 nodes)

• 3 Convolution with 3 × 3 kernel size (3CONV)

• Soft-max out

tu_pian_6.png

Figure 6. AlexNet CNN architectural layout of the proposed technique

This research is built on the suggested AlexNet CNN framework, which is shown in Figure 6. The first picture input layer serves as a preprocessing step for the study project. The spatial dimensions of incoming frames are down-sampled from 640 × 480 to 227 × 227, thereby lowering the processing requirements of the DL architecture. The architecture is made up of five convolutional (CONV) layers, three pooling (POOL) layers, and a rectified linear unit (RELU) after each. A total of 96 kernels with size of 11 × 11 by 3 are used in the first convolutional layer. The second convolutional layer then makes use of 256 kernels that are 5 × 5 in size. 384 kernels with 3 × 3 dimensions are selected for the third, fourth, and fifth layers. An activation map is produced for each convolutional layer. A 2 × 2 stride is used to merge the activation maps (feature maps) from the first, second, and fifth convolutional layers with the 3 × 3 pooling layers. The framework is divided into eight levels, each containing 4096 nodes. These layers are essential for building flexible feature maps that make feature extraction easier. Then, the fully connected layers (FC) get the activation maps, and Soft-max activation is used to generate classification probabilities, which are then used in the final classification step.

Convolution Network Layer

The layer that creates the activation maps that is exposed to classification layers in the DL phenomenon of neural networks is the most important layer. It contains kernel which moves across the source frame (input) to produce the activation map as the output. Everywhere the input, matrix multiplication was done, and the result was then integrated. Eq. (11)'s output activation map is described as:

$N_x^r=\frac{N_x^{r-1}-L_x^r}{S_x^r}+1 ; N_y^r=\frac{N_y^{r-1}-L_y^r}{S_y^r}+1$           （11）

where, $\left(N_x, N_y\right)$ is the width along with the height of activation map that is output for the final layer, $\left(L_x, L_y\right)$, size of kernel and $\left(S_x, S_y\right)$ which determines the amount of pixels omitted by kernel in both vertical and horizontal lines and $r$, index denotes the level or layer i.e., $r=1$. Convolution is performed on the activation map being input and a kernel to produce resultant activation map that is described as:

$X_1(m, n)=(J * R)(m, n)$                  （12）

where, $X_1(m, n)$ is a $2 \mathrm{D}$ output activation map created on combining the 2D kernel $R$ of magnitude $\left(L_x, L_y\right)$ and input activation map $J$ as indicated in Eq. (12). The symbol * denotes the convolution around $J$ and $R$. In Eq. (13), the convolution process is defined below;

$X_1(m, n)=\sum_{p=-\frac{L_x}{2}}^{p=+\frac{L_x}{2}} \sum_{q=\frac{L_y}{2}}^{q=+\frac{L_x}{2}} J(m-p, n-q) R(p, q)$                   （13）

In the proposed structure, the utilization of 5 CONV layers along with RELU layer and response normalization layer for obtaining the highest activation maps generate the input frames needed to train the dataset with optimum ACC.

Rectified Linear Unit Layer

RELU activation technique is applied to each of the versatile layers in the next step in order to make the network non-linear and reinforce it. It effectively compensates regarding non-linear characteristics. It is employed to the map of features produced by the convolutional layer as the output. feature map. With regard to training time, the nonlinear slope descent becomes saturated by the usage of tanh(.) and the RELU activation function. Eq. (14) describes tanh(.) as:

$\begin{gathered}X_2(m, n)=\tanh \left(X_1(m, n)\right)=\frac{\sinh \left(X_1(m, n)\right)}{\cosh \left(X_1(m, n)\right)}=1+ \frac{1-e^{-2 * X_1(m, n)}}{1-e^{-2 * X_1(m, n)}}\end{gathered}$               （14）

where, $X_2(m, n)$ is a $2 \mathrm{D}$ result activation map by performing $\tanh ($.$) to the source activation map X_1(m, n)$ that is acquired after sent via the convolutional layer. The parameters used in the last activation map are acquired using RELU procedure as mentioned below:

$X(m, n)=\left\{\begin{array}{cc}0, & \text { if } X_2(m, n<0) \\ X_2(m, n), & \text { if } X_2(m, n \geq 0\end{array}\right\}$               （15）

where, $X(m, n)$ is derived by adapting the negative numbers to zero and gives the identical number again on obtaining any affirmative result which is demonstrated in Eq. (15). Inclusion this RELU layer into the design that was proposed as deep CNNs obtain substantially faster pace when integrating this specific layer (RELU).

Maximum Pooling Layer

In order to decrease the relative dimension of every frame and the computational cost of the recommended DL structure, a layer is that is added to the suggested framework (pooling layer) despite the primary and next convolution layers and then again following the fifth convolution layer. The pooling method typically averages each slice of the picture or simply selects the highest value. The suggested method employs pooling, which produces superior results on this configuration, by utilizing the maximum value against each slice. Figure 7 shows how to activate the output for down-sampling the pictures while using the maximum pooling layer.

tu_pian_7.png

Figure 7. Maximum pooling layer

Response Normalization Layer and the SoftMax Activation

After the first two sessions, response normalization is carried out to decrease the rate of error on a test set of the recommended system. Along with the data input of the whole network in Eq. (16), this particular layer normalizes the layers of input inside systems. The process of normalizing works as described below:

$N_{e, q}^x=\frac{b_{e, f}^x}{\left(z+\alpha \sum_{j=\max (0, x-c / 2)}^{\min (T-1, x+c / 2)}\left(b_{e, f}^x\right)^2\right)^\gamma}$       （16）

where, $N_x, e, f$ indicates the normalization of the behavior of neurons $b_{e, f}^x$ that is estimated at location $(e, f)$ using the help of the kernel, k. $T$ is the entire spectrum of kernels inside the different layers. $z, c, a, g$ are the consistent values hyperparameters and the values they hold are modified through applying a set of validation parameters correspondingly.

A classifier on the top of the retrieved characteristics is called Soft-max. For classification of multiclass, the output of deep convolutional network layer (DCNN) is supplied to layer Softmax, which aids in calculating classification probabilities, after five sets of convolutional network layer processing. The last classification layer utilizes these probabilities in order to divide each frame into crowd/out-field, close-up, medium and long perspectives.

Dropout Layer

To avoid excessive fitting of the input information by raising the number of iterations by a ratio of two, keeping the neurons densely packed the dropout layer is implemented in the initial pair of layers that are completely connected once the number of iterations twice in the network. It employs neural networks to average models, and it is a particularly effective technique of managing training data. The generated activation map is down sampled to a single frame (pixel) for every map by the processing of maximum pooling layers, convolutional layer kernel sizes, and their skipping factors. The output of the uppermost layers is further related to a 1D feature vector through a layer that is completely connected. In order to extract higher-level characteristics from the training data, the top layer must always be properly connected to the output unit for the class label. This method, a normalization approach on completely linked layers prior to and subsequent dropout is illustrated.

3.4.2 Hyper parameter tuning using hybrid optimization

Ant colony optimization

Ant colony optimization (ACO) is an organically designed MA which imitates the ant's foraging behavior. Since ACO permits parallel processing without creating a process dependence and provides feedback on ant behavior in the search space, the model is more rational than previous MAs [36]. Ants can determine the quickest path between their colony and a food source; therefore, they are not blind while searching for food. Ants drop down a chemical substance known as pheromones along their trails as they go. The pheromone acts as a conduit to facilitate interaction amongst ants and indicates the shortest approach to their food source of food. Ants find food on detecting the pheromones left behind by other ants who have already travelled a certain route, which enhances the likelihood that more ants will follow in their footsteps. ACO bases its probabilistic judgments on the heuristic data and pheromone trail. As they go along a trail, the ants adjust the pheromone amount at any point. A feature has a greater chance of becoming a component of the shortest route the more ants pass over it and the more pheromones are deposited there as a consequence. The most ants will go down the route with the greatest concentration of pheromones, and it will also be the shortest way. Ants have distributed arbitrarily among a group of features corresponding to a predetermined greatest number of several generations $T$ and the pheromone level $t 0=1$ is initiated in each of the features of $M$. The change in probability $T P_i^k(g)$ of the $k t h$ ant at the ith feature in Eq. (17) has shown below [37].

$T P_i^k(g)=\left\{\begin{array}{lr}\frac{\left[\tau_i(g)\right]^\alpha\left[\eta_i\right]^\beta}{\sum_{j \in_{j i}^k}\left[\tau_i(g)\right]^\alpha\left[\eta_i\right]^\beta} & \text { if } j \in j_i^k \\ 0, & \text { otherwise }\end{array}\right\}$         (17)

where, $j_i^k$ is a collection of probable neighbors of $i$ th features that aren't examined using the $k t h$ ant. The relative relevance of pheromone intensity $t i$ and heuristic information $h j$ with regard to ants' actions are given by non-negative constants $a$ and $b$, correspondingly. On selecting an additional feature in the ant's route, a function of fitness (FF) is applied to measure the fresh set of chosen features. The progression of $k t h$ ant is terminated if the increase in the fitness value is not obtained after inserting any additional feature. In Eq. (18), if the halting requirement is not fulfilled, the quantity of pheromone level at succeeding generation $g+1$ at $i t h$ feature is upgraded as

$\tau_i(g+1)=(1-p) \tau_i(g)+\sum_{k=1}^N \Delta \tau_i^k(g)$              (18)

$\Delta \tau_i^k(g)=\left\{\begin{array}{lc}\left.F F\left(S_{k(g)}\right) / \mid S^k(g)\right) \mid, & \text { if } i \in S^k(g) \\ 0, & \text { otherwise }\end{array}\right\}$                    (19)

where, $\mathrm{p}$ serves as the pheromone decaying rates, $0 \leq p \leq 1$ N represents the count of ants, $S^k(g)$ represents number of the chosen features, then Dtik stands for the pheromone dropped by $k t h$ ant if $i t h$ feature is along the simplest route of ants; if not it equals 0; which is mentioned in Eq. (19).

The stoppage criteria are accomplished whenever $g$ hits the preset threshold $T$. The set of features having the greatest pheromone level and least fitness score will be chosen as an OFS.

Reptile search algorithm

To mimic the surrounding and hunting behavior of crocodiles, [25] suggested the reptile search algorithm (RSA) in 2021. It represents a gradient-free technique that may commences with producing the following random responses to Eq. (20):

$x_{i, j}=rand_{\in[0, N]} \times\left(U B_j-L B_j\right)+L B_j$ for $i \in\{1, \ldots, N\}$ and $j \in\{1, \ldots, M\}$      (20)

where, $x_{i, j}$ is its $i$ th response for $j$ th source feature for entire $N$ responses containing $M$ features, $\operatorname{rand}_{\epsilon[0, N]}$ is an arbitrary integer dispersed similarly in the interval $[0,1](0,1)$, then this $j t h$ feature comprises upper $U B_j$ and lower $L B_j$ limits.

RSA may be defined in terms of two principles: exploration and exploitation, much as the other nature-inspired MAs. The crocodile's ability to makeover while surrounding its prey helps to explain these concepts. To benefit from crocodiles' natural behavior, RSA's total iterations are split into four phases. RSA completes the investigation in the first two phases using an encircling habit that includes high and belly walking motions. Crocodiles start to circle the area to explore it, enabling a more thorough search of the solution space. Eq. (21) may be used to quantitatively represent this behavior as follows:

$x_{i, j}(g+1)=\left\{\begin{array}{c}{\left[-n_{i, j(g)} \cdot \gamma \cdot Best_j(g)\right]} \\ -\left[rand_{\epsilon[1, N]} \cdot R_{i, j}(g)\right], g \leq \frac{T}{4} \\ E S(g) \cdot Best _j(g) \cdot x_{\left(rand_{\in[1, N]}, j\right),} \\ g \leq \frac{2 T}{4} \text { and } g>\frac{T}{4}\end{array}\right\}$     (21)

where, $Best_j(g)$ be the optimum response on $j t h$ feature, $n_{i, j}$ pertains to the search operation for the $j t h$ feature in the ith response (estimated using Eq. (22)), the parameter $g$ regulates entire exploration ACC through hout the total number of iterations and is fixed as 0,1 . The reduction component $R_{i, j}$ is employed to decrease the exploration location and is calculated as in Eq. (25), $rand_{\in[1, N]}$ is an integer from 1 and $N$ intended to dynamically choose any of the possible candidate response, and $E S(g)$ an evolutionary sense, that refers to the probability ratio dropping from 2 to -2 across iterations, determined in Eq. (21).

$n_{i, j}=Best_j(g) \times P_{i, j}$                    (22)

where, $P_{i, j}$ denotes the $\%$ distinct differences among the $j t h$ value of the best response and its equivalent amount of value in the present response and is computed in Eq. (23) as:

$P_{i, j}=\theta+\frac{x_{i, j}-M\left(x_i\right)}{ Best _j(g) \times\left(U B_j-L B_j\right)+\epsilon}$             (23)

where, $q$ symbolizes a crucial parameter that governs the exploration effectiveness, $e$ is a small floor value, then $M\left(x_i\right)$ denotes the typical responses in the Eq. (24) which is expressed as:

$M\left(x_i\right)=\frac{1}{n} \sum_{j=1}^n x_{i, j}$                   （24)

$R_{i, j}=\frac{Best_j(g)-x_{{(rand_{\in[1, N]}, j)}}}{Best_j(g)+\in}$          （25）

$E S(g)=2 \times rand_{\in[-1,1]} \times\left(1-\frac{1}{T}\right)$              （26）

where, the measure 2 functions as a multiplier to put forward correlation levels in the interval [0,2], then $\operatorname{rand}_{\epsilon[-1,1]}$ is a randomized numerical quantity within $(-1,1)$ in Eq. (26).

In the conclusive 2 stages, RSA develops the exploitation (hunting) searching feature space to provide optimal solution via the following methods: seeking coordination and cooperation. The solution may modify the values at the time of exploitation utilizing the subsequent Eq. (27):

$x_{i, j}(g+1)=\left\{\begin{array}{c}\operatorname{rand}_{\epsilon[-1,1]} \cdot \text { Best }_j(g) \cdot P_{i, j}(g), \\ g \leq g>\frac{2 T}{4} \\ {\left[\epsilon \cdot \text { Best }_j(g) \cdot n_{i, j}(g)\right]-[(g)],} \\ g \leq T \text { and } g>\frac{3 T}{4}\end{array}\right\}$           (27)

The sophistication of candidate responses at each and every iteration is evaluated by utilizing the predefined FF as well as the algorithm breaks after taking T iteration and a candidate response with the least fitness value is chosen as online forwarding strategy (OFS).