Track quality and train operating safety depend on routine inspections of track components. According to statistics from the Federal Railroad Administration (FRA), train accidents in the United States are primarily caused by the failure of track components. Existing technologies, such as fault detection and fault segmentation, have been used to increase efficiency. However, sometimes these models fail when dealing with small cracks or broken cracks. To address this issue, we suggest implementing a fusion model that uses a deep convolutional neural network for fault detection and semantic segmentation on railway lines. The ResNet50 network is used for fault identification in this model, whereas the Deep Residual UNet network is applied for semantic segmentation. The identified defects are isolated, and decisions can be made from resulting binary image segments on whether to inspect or disregard detected faults. These defects encompass damages observed on railway tracks and any absent or damaged components such as spikes, clips, and rails. Comprehensive tests carried out on a track dataset obtained from the Kaggle repository demonstrate that the suggested approach is capable of accomplishing the following: 1) surpassing current leading models by achieving 94.3% mean average precision (mAP) for fault detection and 79 frames/sec, and 2) achieving 95.6% accuracy in semantic segmentation. Therefore, this fusion model helps inspect tracks by detecting and segmenting faults, thereby reducing derailment cases and other railroad-related accidents.
Image relighting, which involves modifying the lighting conditions while preserving the visual content, is fundamental to computer vision. This study introduced a bi-modal lightweight deep learning model for depth-guided relighting. The model utilizes the Res2Net Squeezed block’s ability to capture long-range dependencies and to enhance feature representation for both the input image and its corresponding depth map. The proposed model adopts an encoder–decoder structure with Res2Net Squeezed blocks integrated at each stage of encoding and decoding. The model was trained and evaluated on the VIDIT dataset, which consists of 300 triplets of images. Each triplet contains the input image, its corresponding depth map, and the relit image under diverse lighting conditions, such as different illuminant angles and color temperatures. The enhanced feature representation and improved information flow within the Res2Net Squeezed blocks enable the model to handle complex lighting variations and generate realistic relit images. The experimental results demonstrated the proposed approach’s effectiveness in relighting accuracy, measured by metrics such as the PSNR, SSIM, and visual quality.
The detection of pesticide residue in vegetables is becoming evident in modernized agriculture. Our present work used visible/near-infrared (400–1000 nm) Hyperspectral Imaging Systems (HSIs) to detect the pesticide residue levels in Brinjal. This paper proposes a deep learning framework named the 3D Squeeze Excitation-Residual Spectral Network (3D SERSN) that takes raw 3-D cubes as input data without feature engineering for hyperspectral image classification. In this network, the spectral residual blocks consecutively learn discriminative features from abundant spectral signatures and the newly introduced SE block inside the residual block keeps the skip connection branch as clean as possible to make the learning identity easy. The proposed network is evaluated on the developed HSI dataset of Brinjals. Image samples were captured using a hyperspectral camera for Brinjals without pesticide and with pesticides at different concentrations of 0.5ml, 3ml and 6ml in 1 litre of water namely low, medium and high. Two types of classification analysis, two classes (with or without pesticide) and four classes (pure, low, medium and high) were done with a set of machine learning algorithms and using the proposed deep network. The classification analysis was also implemented with seven machine learning algorithms and with a CNN. For the four-class classification problem, the Multi-class support vector machine and the Discriminant analysis Algorithm reported the highest accuracy of 89%; for the two-class classification, the K-nearest neighbour classifier recorded the highest accuracy of 78%. The proposed 3D SERSN gives an accuracy of 94% for the two-class problem and 98% for the four-class problem on the proposed dataset for pesticide residue estimation on brinjals, showing the effectiveness of our suggested network for the stated issue.1