3.2. Proposed Method
To reduce background noise, blurriness, and spatial distortion, we took two images of each scene and employed a Multi-Scale Discrete Wavelet Transform (MsclDWT) using hybrid fusion rules method to extract complementary information from the input images. Figure 4 visually shows the MsclDWT method. First and foremost, the input images are decompose to obtain its coefficients using proper fusion rules. To process wavelet coefficients, it is important to applied suitable decomposition scales and wavelet family. Few scales selection causes loss of image information while too many scales selection causes in image blurriness. This study employed a four-level scale for decomposing high frequency components and a two-level scale for decomposing low frequency components, as well as the 'Haar' family. In this study, Principal Component Analysis (PCA) and Consistency Verification functions are employed as a fusion rules. After that, Inverse DWT is applied to get the final fused image. These steps/rules are applied to get complementary information/features. PCA is used in the image fusion process to highlight the silent details in the input images. It is applied to calculate the weights of the coefficients using the following equations.
1) Suppose input modalities coefficients are W1 and W2.
2) Covariance matrix measurements
$$CVN \left({W}^{1},{W}^{2}\right)=E\left[\left({W}^{1}-{\mu }_{1}\right)\left({W}^{2}-{\mu }_{2}\right)\right]$$
2
Where E denotes expectation vector and \({\mu }_{1}, {\mu }_{2}\) are the coefficients which can be calculated using Eq. (3).
$${\mu }_{1}= \frac{1}{n}\sum _{a=1}^{n}{Z}_{i}^{1}$$
3
$${\mu }_{2}= \frac{1}{n}\sum _{a=1}^{n}{Z}_{i}^{2}$$
4
Then Eigen vectors (VCc) and Eigen values (Edv) are calculated using Eq. (5).
$$\left[VCc Edv\right]=eig\left(CVN\right)$$
5
VCc is calculated to obtain normalized weights using equations (6).
\(R{i}_{1}=\frac{VCc\left(1\right)}{\sum \sum VCc}\) , \(R{i}_{2}=\frac{VCc\left(2\right)}{\sum \sum VCc}\) (6)
At the end, fused coefficient can be calculated using Eq. (7).
$${W}_{F}= {W}^{1}\times W{i}_{1}+{W}^{2}\times W{i}_{2}$$
7
And then Consistency Verification is fusion rule is used to reduce errors/wrong pixels values. Here a window size of 9X9 is applied to generate a new mapping window for decision purposes. Following this step, the final image has all complementary information/features.
In the second step, Lab color conversion followed by color thresholding method is employed to detect and segment the normal and rust affected images. Image conversion is a subset of preprocessing that involves converting images into fewer color spaces such as lab space or black and white in order to simplify the computational process. However, conversion is not always necessary; it should it should be avoided when the result of the conversion may affect the spatial and spectral information. Lab color space can be defined mathematically in three axes, where L denotes lightness, “a*” represents the color information of a green-red axis, and “b*” represents the color information for a blue-yellow axis. The important feature of Lab color space is its device independence, which means that colors are defined independently based on the nature of creation. In this study, the RGB values are first converted into a Lab color space model to simplify the computational process and improve the accuracy of rusted pixel detection. The Lab color model comes with a variety of benefits, including device independence and a broad color gamut. Furthermore, it can compensate for the RGB color model's uneven color distribution, which contains an abundance of transition colors from blue to green. Image is converted to lab space for further process of segmentation. The advantage of lab space image is by having 1 channel dedicated to the luminosity of the image and 2 other dedicated to color information. Lab color space is more accurate color space because it allows you to do things you cannot so with RGB color image. It also reduce color which is easy to perform further processing. Figure 5 depicts lab color space converted image.
Finally, Lab color space method is followed by color image thresholding to segment and classified the rusted pixels. The segmentation process is based on different information found in the image. This might be boundaries, color information, and segment of an image. The image background is removed based on whether the image pixel falls above or below the threshold value. It will cause to separate images with a specific end goal to remove the healthy part from wheat leaf, which will break down the infection. We have used color thresholding technique which will convert the unhealthy or rust part of leaf pixel value to 0 and healthy part will be in the same color. It is better for huge measure of information and gives better and accurate result with less time. Equations (8), (9), and (10) have been developed to measure the thresholding values for the yellow color using the minimum and maximum values of the RGB components. The original values, however, have been modified to accommodate a 10% difference [28].
$$gray\left(m,n\right)=\{\frac{f\left(m,n\right), 0\le red\left(m,n\right)\le Trr,}{gray1\left(m,n\right), red\left(m,n\right)>Trr.}$$
8
$$gray\left(m,n\right)=\{\frac{f\left(m,n\right), Trrg\le green\left(m,n\right)\le 1,}{gray1\left(m,n\right), green\left(m,n\right)>Trrg.}$$
9
$$gray\left(m,n\right)=\{\frac{f\left(m,n\right), 0\le blue\left(m,n\right)\le Trrb,}{gray1\left(m,n\right), blue\left(m,n\right)>Trrb.}$$
10
Where gray1(m,n) is the intensity value of pixel and red(m,n), green(m,n), and blue(m,n) are the intensity value for red, green, blue channels. We keep the minimum threshold 0.058 and maximum 99.617 for channel 1, minimum 0 and maximum 14.601 for channel 2, and minimum 0 and maximum 43.442 for channel 3. Then separated background image and converted the segmented image in to binary one that is rust part of the leaf will convert to 1 and background to 0. Moreover, the rust affected area are computed by summing all pixels having intensity value of 1 i.e. rust affected pixels.