Detection and Classification of Night Jasmine Leaf Diseases Using a Novel Deep Learning - Based Optimization Model

3.1 Proposed model

The proposed night jasmine leaf disease detection and classification model is consisted of numerous phases like data collection, pre-processing, segmentation, feature extraction, detection, and classification. The first source of the data describes the Night Jasmine Leaf Database, a Kaggle resource. The bilateral filter, median filter, Wiener filter, as well as CLAHE techniques are used to process the gathered information. The adaptive threshold-based segmentation approach is used to segment this pre-processed information. This segmented data is now subjected to feature extraction using the VGG-16 method. Finally, night jasmine leaf disease is identified and categorized using the new ISSAE method. The optimization technique called PPBO does the usual SSAE parameter modification with the objective function of maximizing return accuracy. The proposed night jasmine leaf disease detection and classification model is diagrammatically depicted in Figure 1.

1.png

Figure 1. Proposed night jasmine leaf disease detection and classification model

3.2 Data collection

The purpose of the comprehensive collection known as the Night Jasmine Leaf Diseases Dataset is to support the study as well as management of leaf diseases that affect the night jasmine plant. The dataset is gathered from the link, “https://www.kaggle.com/datasets/shuvokumarbasak4004/night-jasmine-leaf-diseases-dataset.” High-resolution photos related to various leaf disease stages as well as types are included in this collection, which describes a great resource for plant pathology and agricultural technology researchers and experts. In order to facilitate the development as well as assessment of deep learning methods for illness detection and classification, the photos are arranged to depict different disease situations. The dataset contributes to the advancement of automated plant disease diagnosis by offering a diverse array of visual data, which results in more accurate as well as efficient agricultural practices. It serves as a fundamental tool for developing early disease detection computer vision applications, which helps to lower crop loss as well as promote sustainable agriculture. The dataset is openly available on Kaggle, encouraging collaboration as well as innovation in the field of plant disease management. The dataset comprises symptom-based image labels, as pathogen-level annotations were not available for this study. The proposed framework is thus designed to operate on visual disease cues. The description of the dataset is listed in Table 2.

Table 2. Dataset description

3.3 Pre-processing

Pre-processing enhances the quality as well as consistency of input photos before they are used in deep learning algorithms, which is crucial for the night jasmine leaf disease detection and classification model. By ensuring that the method highlights important patterns related to leaf texture, color variations, as well as disease signs, the first processing stages improve the accuracy of feature extraction and categorization. Effective pre-processing improves the resilience as well as dependability of the detection framework by reducing computing complexity and the chance of overfitting. Here, the pre-processing of the collected night jasmine leaf disease images is done by the bilateral filtering, median filtering, wiener filtering and CLAHE approaches.

Create a two-dimensional matrix representation of the original greyscale picture.

$J(j, k), 1 \leq j \leq N, 1 \leq k \leq 0$                      (1)

Here, $J(j, k)$ is the pixel intensity in an image of size $N \times O$ at the coordinates $(j, k)$.

Bilateral filtering: One non-linear smoothing method that maintains edges describes the bilateral filter, which is described as follows:

$\begin{aligned} & J_c(j, k)=\frac{1}{X(j, k)} \sum_{l=-s}^s \sum_{m=-s}^s H_t(l, m) \cdot H_s(J(j+l, k+m)-J(j, k))\cdot J(j+l, k+m)\end{aligned}$                     (2)

Here, the output after bilateral filtering is shown by $J_c(j, k)$, window radius is shown by $s$, spatial Gaussian kernel is hown by $H_t(l, m)=\exp \left(-\frac{l^2+m^2}{2 \sigma_t^2}\right)$, range Gaussian kernel is shown by $H_s(\Delta)=\exp \left(-\frac{\Delta^2}{2 \sigma_s^2}\right), \Delta= J(j+l, k+m)-J(j, k)$, and the normalization factor is shown by $X(j, k)$ respectively.

$\begin{aligned} X(j, k)=\sum_{l=-s}^s \sum_{m=-s}^s & H_t(l, m)\cdot H_s(J(j+l, k+m)-J(j, k))\end{aligned}$                (3)

This filter combines pixel similarity as well as spatial closeness to reduce noise and preserve edges.

Median filtering: Utilizing a sliding window, the median filter replaces every pixel with the median value of nearby pixels:

$\begin{gathered}J_N(j, k)=\{\text {Median}\}\{J(j+l, k+m) \mid-s\leq l, m \leq s\}\end{gathered}$                (4)

The window's dimensions are $(2 s+1) \times(2 s+1)$, and it is very effective in removing impulse noise, sometimes known as salt-and-pepper noise, where $J_N(j, k)$ describes the filtered output at the pixel at $(j, k)$. The median operator avoids intensity blurring, which protects edges compared to averaging filters.

Wiener filtering: Reducing the mean square error among the estimated as well as real images is the goal of the Wiener filter. In the frequency domain, the formulation is:

$J(v, w)=\left[\frac{|I(v, w)|^2}{|I(v, w)|^2+\frac{T_o(v, w)}{T_g(v, w)}}\right] \cdot \frac{G(v, w)}{I(v, w)}$                    (5)

Here, the Fourier transform related to the degraded image is shown by $G(v, w)$, Wiener-filtered image in frequency domain is shown by $J(v, w)$, Power Spectral Densities (PSDs) associated with the noise as well as the original image is shown by $T_o(v, w)$ and $T_g(v, w)$, and the degradation function like blur kernel is shown by $I(v, w)$ respectively. In the spatial domain (with local window estimates considered):

$J_X(j, k)=\mu+\frac{\sigma^2-\eta^2}{\sigma^2} \cdot(J(j, k)-\mu)$                 (6)

Here, the local variance is shown by $\sigma^2$, local mean in the window is shown by $\mu$, noisy image intensity is shown by $J(j, k)$ and noise variance is shown by $\eta^2$ respectively. When the noise power is known or estimated, this filter is perfect for eliminating Gaussian noise. The image may undergo sequential processing for efficient noise reduction.

$J_{\text {Final }}=X(N(C(J)))$                   (7)

Here, bilateral filtering operation is shown by $C(J)$, median filtering operation is shown by $N(\cdot)$ and wiener filtering operation is shown by $X(\cdot)$ respectively.

CLAHE: An adaptation of the AHE method is CLAHE. CLAHE solves the over amplification issue with normal AHE by making use of the clip limit as well as count of tiles options. The CLAHE method is used to segment the image into $N \times O$ local tiles. Every tile's histogram is calculated separately. In order to compute the histogram, the mean pixel count per region may first be found using Eq. (8).

$O_b=\left(O_y \times O_z\right) / O_h$                   (8)

Here, $O_h$ is the number of grey levels, $O_b$ is the mean number of pixels, $O_y$ is the number of pixels in the $y$ dimension, and $O_z$ is the number of pixels in the $z$ dimension. The clip limit can the be defined as shown by Eq. (9) to restrict the histogram.

$O_{C L}=O_b \times O_{O C L}$                     (9)

Here, $O_{O C L}$ stands for the normalized clip limit, which ranges from 0 to 1 , and $O_{C L}$ stands for the clip limit. The clip limit for the height associated with every tile's histogram is then established using Eq. (10).

$I_j=\left\{\begin{array}{c}O_{C L} \quad \text {if } O_j \geq O_{C L} \\ O_j \quad \text {otherwise}\end{array}\right.$                 (10)

Here, $j=1,2, \cdots, M-1, M$ indicates the count of grey levels, $I_j$ describes the height associated with the $j^{t h}$ tile's histogram, and $O_j$ describes the $j^{t h}$ tile's histogram. The total number of clipped pixels may be determined using Eq. (11):

$O_d=\left(O_y \times O_z\right)-\sum_{j=0}^{M-1} I_j$                  (11)

Here, the count of pixels that are clipped is denoted by O_d. Once O_d has been established, the clipped pixels must be redistributed. Redistribution of pixels can be either uniform or nonuniform. Eq. (12) may be used to calculate the amount of pixels that need to be reallocated.

$O_s=O_d / M$                  (12)

Here, the amount of pixels that need to be reallocated is indicated by O_s. The reduced histogram is then normalized using Eq. (13).

$I_j=\left\{\begin{array}{c}O_{C L} \text { if } O_j+O_s \geq O_{C L} \\ O_j+O_s \quad \text {otherwise}\end{array}\right.$                  (13)

Here, $j=1,2, \cdots, M-1$. To get the number of undistributed pixels, Eqs. (11) and (12) are used. Go over the Eq. (13) again until all of the pixels have been redistributed. In the end, the contextual region's cumulative histogram might be represented using Eq. (14):

$D_j=\frac{1}{O_y \times O_z} \sum_{k=0}^1 I_k$                     (14)

After completing all among the previously described computations, the contextual region's histogram is aligned with Rayleigh, uniform, or exponential probability distributions to attain a desired brightness as well as visual quality. The enhanced image will be the end outcome when the CLAHE processes have been completed.

The use of multiple filters in the preprocessing stage is motivated by the heterogeneous noise characteristics observed in night jasmine leaf images. Bilateral filtering preserves edges while removing Gaussian noise, median filtering effectively mitigates impulse noise, Wiener filtering suppresses blur-induced distortions, and CLAHE enhances contrast under uneven illumination. Since no single method is capable of simultaneously addressing all these distortions, the combined pipeline provides a balanced enhancement strategy that improves the quality of downstream feature extraction.

3.4 Segmentation by adaptive threshold-based segmentation

By isolating the afflicted leaf regions from the background, segmentation is essential to the night jasmine leaf disease detection and classification model because it enables the method to concentrate only on the relevant portions of the picture. By separating the leaf region from dirt, shadows, or various adjacent objects, segmentation reduces noise as well as enhances the delineation of disease-associated characteristics, such as spots, discoloration, or texture changes. This focused isolation improves classification accuracy by enabling the algorithm to pinpoint particular characteristics linked to illness severity as well as trends. Furthermore, segmentation improves resource effectiveness by handling less irrelevant information, which lowers computing needs. By ensuring that forecasts only consider the affected leaf areas, it further improves the interpretability of the method as well as makes the disease detection framework more robust, reliable, and effective in real-world applications. Here, the segmentation process for the pre-processed image of the night jasmine leaf disease model is accomplished by the adaptive threshold-based segmentation approach.

In order to effectively distinguish among foreground as well as background, adaptive threshold-based segmentation describes a dynamic method for picture segmentation that adjusts the threshold value locally depending on the pixel neighborhood, even in the presence of irregular lighting or contrast variations. Instead of relying on a single overall threshold, this method separates a picture mathematically by computing a locally varying threshold for each pixel. Assume $J(j, k)$ be the representation of the greyscale picture input, where $(j, k)$ are the spatial coordinates associated with the $N \times O$ image. For every pixel $J(j, k)$, a square window with dimensions $x \times x$ is considered. The local standard deviation $\sigma_x(j, k)$ as well as local average $\mu_x(j, k)$ are calculated at this time. Next, the adaptive threshold $U(j, k)$ is computed as follows:

$U(j, k)=\mu_x(j, k)+l \cdot \sigma_x(j, k)$                    (15)

Here, $l$ describes a constant that affects how sensitive the threshold is to changes in local intensity. Segmentation is used to construct the binary image $T(j, k)$ utilizing

$T(j, k)=\left\{\begin{array}{cc}1 & \text {if J}(j, k) \geq U(j, k) \\ 0 & \text {otherwise}\end{array}\right.$                  (16)

This technique adapts to local picture properties as well as works particularly well for textured sceneries, medical photos, and documents when uniform thresholding is insufficient. To increase resilience in some variations, the arithmetic mean may be swapped out for the local median or a weighted average using a Gaussian kernel. Additionally, the window size w is crucial; a bigger window produces smoother transitions, while a smaller window provides higher sensitivity to local variations. Adaptive thresholding uses contextual intensity distributions to provide a mathematically efficient method for achieving precise segmentation in complex visual environments.

3.5 Feature extraction by VGG-16

The night jasmine leaf disease detection and classification model relies heavily on feature extraction, which transforms unprocessed picture data into meaningful descriptions that highlight the unique characteristics of both healthy as well as diseased leaves. By focusing only on the major relevant data, effective feature extraction reduces dataset complexity, increases method generalization, as well as enhances classification accuracy. It ensures that the classification framework can more accurately distinguish between distinct disease kinds or healthy leaves, which will lead to a more reliable as well as efficient disease detection procedure.Here, the features are extracted from the segmented images of the proposed night jasmine leaf disease model by the VGG-16 approach.

A deep CNN (Convolutional Neural Network) represents the VGG-16. It has three fully connected layers, five max-pooling layers, thirteen convolutional layers, as well as a Softmax output layer. The Conv Layer's main tasks include making local connections, employing neurons as filters, as well as applying convolution operations on local input via applying filters and a sliding window technique. The generalized formula associated with the operating procedure is:

$B(j, k)=(Y * G)(j, k)+c$                   (17)

Here, * describes the convolution operator, $Y$ describes the input matrix, $c$ describes the bias, and $B(j, k)$ describes the position value related to filter matrix $G$ with respect to output matrix $(j, k)$.

The order $y$ associated with the final output matrix is as follows if the input matrix has order $o$, the filter matrix is of order $g$, and border padding as well as stride are involved:

$y=\frac{o+2 q-g}{t}+1$                      (18)

The activation function in the ReLU Layer enhances the nonlinearity of the method by applying a nonlinear mapping to the feature matrix produced from the convolutional Layer. Rectified Linear Units, or ReLUs, are widely used and are among the major significant unsaturated activation functions.

The main purpose of the pooling layer, also known as down sampling, is to reduce the number of parameters as well as their dimensionality. The idea is to divide the feature graph into many non-overlapping regions using different Windows, then apply average pooling or maximum pooling on these regions.

The data properties are obtained in the Fully Connected Layer after the convolution as well as pooling processes of the convolution and pooling layers, respectively. The correlation among the fully connected layer as well as the outputs related to linear neurons is shown as follows:

$g(y)=\sum\left(x_{j k} y_j\right)+c_j$                    (19)

Here, b is the bias component in the fully connected layer's output.

3.6 Detection and classification by novel ISSAE

The core components related to the night jasmine leaf disease model are detection and classification, which enable the framework to automatically identify as well as differentiate infected leaves from healthy ones. The detection stage focusses on locating the diseased leaf sections, sometimes using segmentation findings or bounding box approaches to pinpoint the exact places displaying symptoms like blights, spots, or color changes. Following the identification of the afflicted regions, the classification stage assigns the leaf to a specific category, such as healthy, mildly infected, or severely infected, or, in the event that many disease types are found, identifies the exact kind of illness. Together, detection as well as classification forms a robust procedure that not only identifies diseased leaves yet also informs farmers about the kind and severity associated with the disease, enabling them to make more accurate and timely agricultural decisions.The detection and classification of the extracted features of the proposed night jasmine leaf disease model is done here by the novel ISSAE approach, where the parameters of SSAE are tweaked by PPBO with the intention of maximizing accuracy as the fitness function.

There are several advantages of using the SSAE to detect as well as classify illnesses in night jasmine leaves. First of all, SSAE excels in learning compressed as well as hierarchical descriptions associated with input pictures, spotting subtleties that conventional techniques would overlook, such as texture variations, color patterns, and disease signs. By ensuring that only the important as well as instructive neurons are triggered, the sparsity requirement reduces noise and improves the method's capacity for generalization. Even in situations where labelled data is few, the framework may improve classification accuracy by SSAE to obtain more abstract as well as differentiating characteristics. Additionally, SSAE can efficiently handle high-dimensional picture information, reducing processing burden while maintaining important details. As a result, the method can withstand variations in lighting, backdrop as well as leaf placement. Using SSAE improves the detection as well as classification procedure, enabling quick and precise identification of night jasmine leaf diseases and assisting farmers in taking prompt action to protect their crops.

The SSAE has several limitations even if it offers strong feature learning for identifying as well as categorizing illnesses of night jasmine leaves. Its high computational complexity describes a major drawback, especially when working with large picture datasets or more complicated network topologies, which leads to lengthy training times as well as a substantial hardware need. Furthermore, without extensive testing, SSAE tuning becomes more difficult due to the influence of hyper parameter settings, such as the learning rate, number of hidden units, as well as sparsity penalties. When the complexity related to the method surpasses the amount of labelled information that is present, overfitting may occur, especially in agricultural datasets, in which annotated pictures are often limited. Furthermore, SSAE training typically entails unsupervised pre-training and fine-tuning, adding more steps than end-to-end methods like CNNs. Unless carefully optimized, these limitations may hinder real-world implementation and make SSAEs less suitable for real-time or resource-constrained applications.

Compared to the traditional SSAE, the ISSAE provides a number of advantages for identifying as well as categorizing illnesses of night jasmine leaves. By using attention mechanisms, optimization strategies or advanced regularization approaches, ISSAE enhances feature learning and is able to recognize more unique as well as disease-specific patterns from leaf pictures. This results in increased resistance to noise, background fluctuations as well as illumination variations. Using optimized architectures or adaptive learning approaches, ISSAE often improves convergence speed as well as reduces training time, making it more efficient on large datasets. Additionally, its improved sparsity management ensures that only the major important characteristics are emphasized, supporting the method's ability to generalize even with little labelled information. Additionally, the improved architecture helps to lessen overfitting, which is a common issue with deep methods, especially when it comes to agricultural datasets. ISSAE is highly successful for early as well as precise disease detection in night jasmine leaves due to its greater accuracy, improved stability, and enhanced computing effectiveness.

Many sparse auto-encoders are set up to form SSAE, a deep network format. The output associated with the hidden layer z_m for each hidden layer m in the group $\{1, \cdots, M-1\}$ related to the SSAE method is described as below:

$z_m=g\left(c_i i_{M-1}+c_m\right)$                   (20)

Here, $M$ is the total number of layers, $g$ describes the dimensionality related to latent factor, $c$ describes the bias, $X$ describes the weight and the hidden description is shown by $i$ respectively. The first $M / 2$ layers associated with the method are the encoder, and the latter $M / 2$ layers are the decoder. It is suggested that the M/2 layer in the SSAE method should generate the latent component, and that there should only be one hidden layer close to it. To replicate the input as well as lower the squared loss among the produced outputs, the SSAE uses a deep framework. The below formula is used to determine the loss function for SSAE.

$K_{\text {Spare}}\left(X_m, c_m\right)=K\left(X_m, c_m\right)+\beta \sum_{j=1}^n L M\left(\rho \| \hat{\rho}_j\right)$    (21)

Here, the weight matrix is shown by $X_m$, bias vector associated with every layer is shown by $c_m$, divergence is shown by $L M$, loss function of SAE is shown by $K_{\text {Spare}}$, weight employed to control sparsity penalty factor is shown by $\beta$ and $\sum_{j=1}^n L M\left(\rho \| \widehat{\rho}_j\right)$ is shown as below.

$\sum_{j=1}^n L M\left(\rho \| \hat{\rho}_j\right)=\sum_{j=1}^n \rho \log \frac{\rho}{\hat{\rho}_j}+(1-\rho) \log \frac{1-\rho}{1-\hat{\rho}_j}$                (22)

Here, $\rho=\frac{1}{n} \sum_{j=1}^n\left(b_k\left(y_j\right)\right)$ describes the average activation of entire training instances in hidden layer neuron $k$, activation value in hidden layer neurons is shown by $b_k$ and the count of hidden units per layer is shown by $n$ respectively. The proposed novel ISSAE for the detection and classification of night jasmine leaf disease is diagrammatically shown in Figure 2.

2.png

Figure 2. Proposed ISSAE for the detection and classification of night jasmine leaf disease model

The ISSAE architecture differs from the traditional SSAE through three key enhancements. First, an adaptive sparsity regulation mechanism dynamically adjusts the sparsity penalty based on reconstruction behavior, preventing neuron under-utilization or collapse. Second, ISSAE employs progressive layer-wise fine-tuning, which stabilizes convergence by refining each hidden layer sequentially instead of training the entire stack at once. Third, PPBO-assisted hyperparameter stabilization ensures optimal selection of sparsity coefficients, hidden units, and learning rates, reducing sensitivity to initialization. Together, these enhancements enable ISSAE to extract more discriminative and stable representations from leaf images compared to the conventional SSAE.

3.7 PPBO algorithm

By fine-tuning the method parameters, optimization is crucial to increasing the efficacy related to the night jasmine leaf disease detection and classification model. In order to improve stability, model convergence, as well as generalization, it helps select the best hyperparameters (like batch size, learning rate, and count of hidden layers). The resilience, effectiveness, as well as capacity of the pipeline to produce precise illness forecasts under various circumstances are all improved via optimization.Here, the PPBO algorithm tunes the parameters of the SSAE model in the proposed night jasmine leaf disease detection and classification model in order to return the accuracy maximization as the objective function.

The process of publishing papers is replicated by PPBO. The main driving force behind PPBO is writers' efforts to improve their work in response to reviewers' as well as editors' comments in order to get acceptance. After outlining the PPBO format, its mathematical modelling is developed by simulating the publication procedure of papers.