Welding is an extensively used technique in manufacturing, and as for every other process, there is the potential for defects in the weld joint that could be catastrophic to the manufactured products. Different welding processes use different parameter settings, which greatly impact the quality of the final welded products. The focus of research in weld defect detection is to develop a non-destructive testing method for weld quality assessment based on observing the weld with an RGB camera. Deep learning techniques have been widely used in the domain of weld defect detection in recent times, but the majority of them use, for example, X-ray images. An RGB image-based solution is attractive, as RGB cameras are comparatively inexpensive compared to X-ray image solutions. However, the number of publicly available RGB image datasets for weld defect detection is comparatively lower than that of X-ray image datasets. This work achieves a complete weld quality assessment involving lap shear strength prediction and visual weld defect detection from an extremely limited dataset. First, a multimodal dataset is generated by the fusion of image data features extracted using a convolutional autoencoder (CAE) designed in this experiment and input parameter settings data. The fusion of the dataset reduced lap shear strength (LSS) prediction errors by 34% compared to prediction errors using only input parameter settings data. This is a promising result, considering the extremely small dataset size. This work also achieves visual weld defect detection on the same limited dataset with the help of an ultrasonic weld defect dataset generated using offline and online data augmentation. The weld defect detection achieves an accuracy of 74%, again a promising result that meets standard requirements. The combination of lap shear strength prediction and visual defect detection leads to a complete inspection to avoid premature failure of the ultrasonic weld joints. The weld defect detection was compared against the publicly available image dataset for surface defect detection.

Deep learning based visual cognition has greatly improved the accuracy of defect detection, reducing processing times and increasing product throughput across a variety of manufacturing use cases. There is however a continuing need for rigorous procedures to dynamically update model-based detection methods that use sequential streaming during the training phase. This paper reviews how new process, training or validation information is rigorously incorporated in real time when detection exceptions arise during inspection. In particular, consideration is given to how new tasks, classes or decision pathways are added to existing models or datasets in a controlled fashion. An analysis of studies from the incremental learning literature is presented, where the emphasis is on the mitigation of process complexity challenges such as, catastrophic forgetting. Further, practical implementation issues that are known to affect the complexity of deep learning model architecture, including memory allocation for incoming sequential data or incremental learning accuracy, is considered. The paper highlights case study results and methods that have been used to successfully mitigate such real-time manufacturing challenges.

This work demonstrates the use of Deep Learning-based Computer Vision for Remanufacturing end-of-life consumer electronics products, considering smartphones as the use-case. We implemented automated detection of screen defects such as scratches and cracks. In turn, this could lead to increased reuse of smartphones in a secondary market alongside new ones to meet consumer demand. The refurbishment of smartphone devices is a growing industry heavily dependent on manual labor, making decisions subjective, especially in grading the severity of damage. A critical analysis of defect detection and smartphone grading from a remanufacturing perspective is conducted. This analysis is significant in a remanufacturing sector that deals with dynamically varying input of cores (used products for remanufacturing), characteristics, timing, and number of returns. The solution we propose here is novel in its own right, requiring data analysis and collection, data curing, defect parameterization, and dataset building to enable model-based training and detection experiments. We collected and annotated a dataset to detect and grade the various defects based on their severity. A range of deep learning models was trained on the dataset to obtain baseline results for the state-of-the-art deep learning detection models, including YOLOv7, YOLOv8, YOLO11 variants, and Mask R-CNN. Our experiments also showed improved precision values when the pre-trained models were pre-fine-tuned using a road crack segmentation dataset before training on our phone defect dataset. The inference time for the YOLOv8x model is 8ms. This reduced inference time with a high precision of 70.4% indicates that a consistent, fast, and accurate grading is achieved here, ensuring a high throughput rate in the remanufacturing process and ensuring sustainability.

The automotive industry has made significant strides in enhancing road safety and enabling automated driving features through advanced computer vision techniques. This is particularly true for short-range vehicle automation, where non-linear fisheye cameras are commonly used. However, these cameras are challenged by optical distortions, known as fisheye geometric distortions, which lead to object deformation within the image and significant pixel distortion, particularly at the image periphery. Based on the observation that fisheye and spherical images exhibit at least superficially similar geometric characteristics, we investigate the applicability of spherical models—including Spherical Convolutional Neural Networks (CNNs) and Spherical Vision Transformers (ViTs)—to fisheye images, even though fisheye images are not truly spherical. We perform our comparison using fisheye datasets—Woodscape, SynWoodscape, and SynCityscapes in autonomous driving scenarios, with a specific focus on the ability of spherical methods (Spherical CNNs and ViTs) to manage fisheye distortions and compared them against traditional non-spherical methods. Our findings indicate that spherical methods effectively address fisheye distortions without needing extra data augmentations. This results in better mean Intersection over Union (mIoU) scores, pixel accuracy, and better surround-view perception than other modern approaches for fisheye semantic segmentation. However, we also find that spherical methods have a greater tendency to overfit smaller datasets compared with non-spherical models. These advancements highlight how non-linear camera images can take advantage of spherical approximations through spherical models in autonomous driving.

Reliable perception in automated vehicles under adverse conditions, such as fog, rain, snow, and lens defocus, is essential for maintaining the safety of road actors and particularly of vulnerable road users. While prior work has primarily focused on camera occlusions, the impact on RADAR and LiDAR remains underexplored, particularly in a unified Bird's Eye View (BEV) space. To address this gap, we first apply occlusion to all three primary sensors: camera, RADAR, and LiDAR, and then systematically investigate its impact by projecting their outputs into the Bird's Eye View (BEV) space for unified analysis of vehicle and map segmentation. A parametrised occlusion pipeline is developed to apply occlusions to each of the sensor modalities. We evaluate both geometry-based and transformer-based fusion architectures, revealing that transformer-based architectures consistently demonstrate greater robustness to sensor degradation. Notably, we demonstrate that BEVCar achieves 45.6% vehicle Intersection-over-Union (IoU) and 53.6% Mean Intersection-over-Union (mIoU) under camera occlusion, surpassing other State-of-the-art (SOTA) models such as MMTraP (37.9% IoU / 47.9% mIoU) and CVT (36.0% IoU / 46.6% mIoU). These improvements are statistically significant (paired t-tests with 95% CI bootstrap, $p < 0.001$). Furthermore, projecting camera features into the BEV space using a backward projection strategy seems to offer greater resilience to occlusion than forward projection. These insights highlight the importance of architectural design, projection choice, and multi-sensor fusion in developing robust perception systems for automated driving under realistic multi-sensor occlusions.

Vision-Language Models (VLMs) are becoming increasingly powerful, demonstrating strong performance on tasks that require both visual and textual understanding. Their strong generalisation abilities make them a promising component for automated driving systems, which must handle unexpected corner cases. However, to be trusted in such safety-critical applications, a model must first possess a reliable perception system. Since critical objects and agents in traffic scenes are often at a distance, we require systems that are not “shortsighted,” i.e., systems with strong perception capabilities at both close (up to 20 meters) and long (30+ meters) range. With this in mind, we introduce Distance-Annotated Traffic Perception Question Answering (DTPQA), the first Visual Question Answering (VQA) benchmark focused solely on perception-based questions in traffic scenes, enriched with distance annotations. By excluding questions that require reasoning, we ensure that model performance reflects perception capabilities alone. Since automated driving hardware has limited processing power and cannot support large VLMs, our study centers on smaller VLMs. We evaluate several state-of-the-art (SOTA) small VLMs on DTPQA and show that, despite the simplicity of the questions, these models significantly underperform compared to humans (∼60% average accuracy for the best-performing small VLM versus ∼85% human performance). However, the human sample size was relatively small, which imposes statistical limitations. We also identify specific perception tasks, such as distinguishing left from right, that remain particularly challenging. We hope our findings will encourage further research into improving the perception capabilities of small VLMs in traffic scenarios, making them more suitable for automated driving applications.