The overall block diagram of the proposed method is shown in Fig. 1. Bone sanding technique is used to convert the normal bone into osteoporosis bone. To execute image denoising, the acquired RGB image is turned into grayscale image to minimize colour complexity. Butterworth low-pass filter is used to remove the unwanted signals and noise to produce smoothened image as it attenuates the high frequency and eliminates the low frequency. Regression in machine learning is used to compare the estimated value between the normal bone and osteoporosis bone. Thus the osteoporosis brittleness of the bone is predicted.
A. BONE SANDING
Sanding method is used to smoothed or polish the surface with sandpaper or mechanical sander. Here, we used this method to reduce the bone thickness thereby converting the normal bone into osteoporosis bone.
B. THERMAL RGB IMAGE
After bone sanding, thermal RGB images of normal bone and osteoporosis bone are captured using thermal camera and viewed using thermal Smart View software which is shown in Fig. 2 and Fig. 3Thermal imaging might potentially be used to improve visibility of things in low-light situations by detecting infrared and generating a picture based on that data.
C. BUTTERWORTH LOWPASS FILTERING
Butterworth low-pass filter is used to remove noise, as low-pass filter attenuates high frequency by eliminating low frequency to create the blurred or smoothened image. For image processing application, processing of colored image makes it complicated. So, RGB image is converted into Grayscale image to reduce color complexity and simplify mathematics. The magnitude spectrum shows variation in pixel values and tells you how powerful the harmonics are in an image. The phase spectrum depicts the location of this harmonic in space as well as variation in intensity.
In Fourier spectrum the transform's origin is shifted to the middle of the frequency rectangle.
NON-DECIMATED WAVELET TRANSFORM
Non decimated wavelet transform algorithm is used to improve the resolution of the edges to make it better for further operations. The wavelet transform permits extracting data of stationary and non-stationary signal variations in time and frequency  i.e., distinctive their frequency of occurrence, localization in time, and creating a reliable approximation of magnitude of this variation. It is possible to present an original grey image, a noisy image, a denoised image based on wavelet, and a denoised image based on stationary wavelet..
D. REGRESSION IN MACHINE LEARNING
Regression is the task of predicting continuous quantity. The analysis is carried out on the proposed Simulink supported ANN. For training the neural network we had Five hidden layers and one output layer. Various factors like gradient, Mu, epochs, time, performance and validation checks are used to proceed. Training state, Performance, error histogram, and regression are all included in the plot section.
The algorithms used are:
a. Random (Data division)
b. Levenberg-Marquardt (Training)
c. Mean Square Error (Performance)
d. MEX (Calculations)
A sample snapshot of neural network training displaying various parameters is shown in Fig. 4.
E. RESULT AND DISCUSSION
The best validation performance is 5.8971e-09 at epoch 350 as shown in Fig. 5 and 0.0086931 at epoch 82 is shown in Fig. 6 for normal & OP bones respectively. The performance is mse against number of epochs. The graph depicts train, validation, and test trends.
The respective values of gradient, Mu and validation check at epoch 413 are 4.4649e-07, 1e-07 and 0 for normal bone shown in Fig. 7 and at epoch 88 are 0.00016411, 1e-08 and 6 for OP bone respectively is shown in Fig. 8. The plot is number of epochs against training state parameters.
The error histogram plot is errors against instances. The error is nothing but outputs subtracted from targets. Here, the error histogram is represented with 20 bins. Bin is the number of vertical bar observed in the graph. The plot shown in Fig. 9 and Fig. 10 shows trends of training, validation, test and zero error.
For normal bone, zero error is obtained at the 16th bin at -5.9e-05 and for OP bone, the zero error is obtained in the 12th bin at 0.01227.
The regression plots for training, validating and testing along with combination of their three plots (all) are shown in the Fig. 11 & 12. The plot displays the output with respect to targets of training, validating and testing. In normal bone regression plot, the data (pixels) is perfectly fitted along the dash line, where the output is equal to the target. The plots are plotted with value of R equal to 1 obtained by comparing the normal bone with normal bone. This case is same when OP bone is compared with OP bone. From this it is clear that the regression plot does not vary and fit perfectly in the dash line when same type of bone is trained.
But in OP bone, for perfect fit the data should fall along the dash line of 45 degree. The plots are plotted with value of R greater than 0.87027 i.e. <1 obtained by comparing normal bone with OP bone indicating that the bone is brittle. When the value of R obtained is less than 1, it means that the bone is bone is brittle and thus the disease osteoporosis can be detected.