Autism spectrum disorder (ASD) is a neurological issue that impacts brain function at an earlier stage. The autistic person realizes several complexities in communication or social interaction. ASD detection from face images is complicated in the field of computer vision. In this paper, a hybrid GEfficient-Net with a Gray-Wolf (GWO) optimization algorithm for detecting ASD from facial images is proposed. The proposed approach combines the advantages of both EfficientNet and GoogleNet. Initially, the face image from the dataset is pre-processed, and the facial features are extracted with the VGG-16 feature extraction technique. It extracts the most discriminative features by learning the representation of each network layer. The hyperparameters of GoogleNet are optimally selected with the GWO algorithm. The proposed approach is uniformly scaled in all directions to enhance performance. The proposed approach is implemented with the Autistic children’s face image dataset, and the performance is computed in terms of accuracy, sensitivity, specificity, G-mean, etc. Moreover, the proposed approach improves the accuracy to 0.9654 and minimizes the error rate to 0.0512. The experimental outcomes demonstrate the proposed ASD diagnosis has achieved better performance.

The redundancy of a multichannel data transmission system increases its reliability. During the operation of the system, it is necessary to diagnose and switch failed channels. To solve this problem, the set of input signals of the system is considered as a vector signal, whose scalar components are the coordinates of the vector in a given dimensional space. The diagnosis is performed using a scalar criterion, whose relative simplicity is ensured by the linearity of the signal transformations applied. To minimize the total probability of diagnostic error, the task of optimizing the tolerance on the diagnostic parameter is solved. The possibility of technical implementation of the proposed method is shown based on matrix transformations of the system's input and output signals. The system efficiency was assessed according to the "reliability-cost" criterion. Scientific novelty of the work consists in the fact that analytical expressions for matrix transformations of input and output vector signals of a multichannel data transmission system have been developed. Realization of these transformations provides diagnostics of the system according to the developed scalar criterion both in the test mode and in the mode of functioning as intended. 

In recent times People commonly display substantial intra-class variability in both appearance and position, making pedestrian recognition difficult. Current computer vision techniques like object identification as well as object classification has given deep learning (DL) models a lot of attention and this application is based on supervised learning, which necessitates labels. Multimodal imaging enables examining more than one molecule at a time, so that cellular events may be examined simultaneously or the progression of these events can be followed in real-time. Purpose of this study is to propose and construct a hybrid machine learning (ML) pedestrian identification model based on multimodal datasets. For pedestrian detection, the input is gathered as multimodal pictures, which are then processed for noise reduction, smoothing, and normalization. Then, the improved picture was categorized using metaheuristic salp cross-modal swarm optimization and optimized using naive spatio kernelized extreme convolutional transfer learning. We thoroughly evaluated the proposed approach on three benchmark datasets for multimodal pedestrian identification that are made accessible to the general public. For several multimodal image-based pedestrian datasets, experimental analysis is done in terms of average precision, log-average miss rate, accuracy, F1 score, and equal error rate. The findings of the studies show that our method is capable of performing cutting-edge detection on open datasets. proposed technique attained average precision of 95%, log-average miss rate of 81%, accuracy of 61%, F1 score of 51%, equal error rate of 59%.

Fracture surface analysis is crucial in investigating manufacturing failures and material characterization. Traditional manual inspection methods are slow and subjective, prompting the need for efficient automated tools using advanced computer vision techniques. Recent machine learning models for classifying surface fractures show potentials but struggle due to the lack of large, labeled datasets. This study explores the potential application of autoencoders, a self-supervised neural network, to identify unintended fracture surfaces from anomalous manufacturing of tungsten-heavy alloys. The proposed autoencoder-based model achieves 97% accuracy in distinguishing undesirable fracture patterns by analyzing the reconstruction loss of the images, surpassing existing methods. This high accuracy highlights the autoencoder's ability to automatically extract and reduce dimensional features from fracture surfaces effectively. The experimental result obtained on tungsten-heavy alloys demonstrate the model's potential towards developing autoencoder-based automated tools for fractographic analyses across various materials and operational scenarios. 

This study explores hyperspectral image classification, particularly focusing on spectral unmixing techniques applied to the widely used "PaviaU" dataset. Nine distinct endmembers, representing materials such as Water, Trees, and Shadows, serve as the foundation for our investigation. Introducing a novel linear regression model meticulously tailored for hyperspectral image reconstruction, we aim to address the complexities inherent in such datasets. Our approach leverages a fusion of non-negative least squares (NNLS) and a sum-to-one constraint, employing the Sequential Least Squares Quadratic Programming (SLSQP) method to seek optimal coefficients.  Through rigorous experimentation and analysis, our model achieves a mean reconstruction error of 1152.318. The efficacy of our approach lies in its seamless integration of NNLS and SLSQP, customizing a solution to the intricate nuances of hyperspectral data. By significantly reducing reconstruction errors, our method represents a substantial advancement in spectral unmixing techniques. Furthermore, this study produces nine abundance maps for each endmember using least squares with constraints, lasso, and averaging the squared differences between observed and reconstructed spectra for pixels with nonzero class labels to determine reconstruction error. Emphasizing the importance of abundance maps and reconstruction errors, we compare the results obtained through our proposed spectral unmixing methods with those of alternative approaches. This comprehensive analysis not only sheds light on the performance superiority of our proposed methods but also provides valuable insights for practitioners and researchers working with hyperspectral imaging data. By offering enhanced accuracy and efficiency in spectral unmixing, our approach holds significant promise for applications ranging from environmental monitoring to precision agriculture and beyond.

New area of image processing termed "digital image forensics" aims to gather quantifiable proof of a digital image's authenticity and place of origin. Detection of forgery images to look for copied and pasted portions; however, depending on whether the copied portion underwent post-processing before being transferred to another party, the detection method may differ. Zernike Moments and Scale-Invariant Feature Transform (SIFT) combined are unique techniques that aid in the identification of textured and smooth regions. But compared to SIFT separately, this combination is the slowest. So in the proposed work, Block based image division and SIFT based key point detection model is developed to detect forgery images. The gathered images are poor visual quality and various dimension, so it is resized and converter grayscale conversion. In addition, pixel values of images are improved using optimal Gaussian filter and adaptive histogram equalization which remove noise and blurring based on sigma value. Then, using the SIFT key point extraction algorithm to extract the image's key point and compute the feature vector of each key-points. In that using a block based matching technique to split the pre-images into blocks, and each blocks are diagonally subdivide. Length of the feature vector is computed using Zernike moments of each blocks. Both SIFT features and Zernike moments features are matched to identify the manipulated image from the given data. The proposed model provides 100% recall, 98.2% precision, and 99.09% F1_score. Thus provide the proposed model was effectively detects forgery image in the given data. 

Non-Small Cell Lung Cancer (NSCLC) represents a significant health challenge globally, with high mortality rates largely attributed to late-stage diagnosis. This paper details a novel approach for denoising computed tomography (CT) scans through 2-dimensional Fractional Fourier transform (2D-FrFT), which has been reported to be effective for time-frequency signal/image processing applications. To establish a foundation for the FrFT filtering of the original and corrupt dataset, a variable fractional-order image processing technique was used. Based on the derived pre-processing of CT scans, a classification model was developed with hand-crafted features and a 2-layer neural network to classify 4834 CT scans collected from the Lung Image Database Consortium image (LIDC-IDRI) dataset into classes of normal lungs and NSCLC infected lungs. This work presents an approach to improving the performance of NSCLC detection through a lightweight neural network that attains 1.00 accuracy, 1.00 sensitivity, and 1.00 AUC. An additional real-time lung cancer dataset from PGI Rohtak, Haryana, has been considered to validate the model and prove its performance against overfitting.  The experimental analysis showed better results than the existing methods for both LIDC-IDRI and hospital datasets and could be a competent assistant to clinicians in detecting NSCLC.

This project aims to enhance online experiences quality by giving users greater control over the content they encounter daily. The proposed solution is particularly valuable for parents seeking to safeguard their children, educational institutions striving to foster a more conducive learning environment, and individuals prioritising ethical internet usage. It also supports users who wish to limit their exposure to misinformation, including fake news, propaganda, and disinformation. Through the implementation of a browser extension, this system will contribute to a safer internet, reducing users' vulnerability to harmful content and promoting a more positive and productive online environment. The primary objective of this work is to develop a browser extension that automatically detects and censors inappropriate text and images on web pages using artificial intelligence (AI) technologies. The extension will enable users to personalise censorship settings, including the ability to define custom prohibited words and toggle the filtering of text and images. Accuracy estimates for various classifiers such as Random Forest (0.879), Logistic Regression (0.904), Decision Tree (0.878), Naive Bayes (0.315), and KNN (0.832) were performed.