Object Recognition (Computer Vision)

Video event detection (VED) is a challenging task especially with a large variety of objects in the environment. Even though there exist numerous algorithms for event detection, most of them are unsuitable for a typical consumer purpose. A hybrid method for detecting and identifying the moving objects by their color and spatial information is presented in this paper. In tracking multiple moving objects, the system makes use of motion of changed regions. In this approach, first, the object detector will look for the existence of objects that have already been registered. Then the control is passed on to an event detector which will wait for an event to happen which can be object placement or object removal. The object detector becomes active only if any event is detected. Simple training procedure using a single color camera in HSV color space makes it a consumer application. The proposed model has proved to be robust in various indoor environments and different types of background scenes. The experimental results prove the feasibility of the proposed method.

This paper presents a combination of techniques which recognize and categorize observed rigid objects, and compute their 3D pose with respect to an industrial robot's workspace frame. This allows an industrial robot to simultaneously recognize and follow the displacements of a specific rigid object, among other objects that may coexist in the same environment.

A research area in Computer Vision focuses on the identification of articulated objects, such as human actions and movements of the hand, which can be used in human-computer interaction, surveillance, and other tracking systems. Two problems arise: identify when two articulated objects in different stances are in the same class of objects, and differentiate the distinct positions of the same object. In both cases, it is necessary to know how correspond the different points or regions of such objects standing in different attitudes. This article presents the Contour-Point Signature; a point descriptor that allows to establish a method to achieve the better matching of points between two figures, and to thus obtain a transformation which relates them. With this descriptor, we can achieve more accurate shape features and implement more efficient retrieval under multi-resolution. In addition, CPS is robust to rigid translation, scaling, rotation and independent of the origin point. A measure of dissimilarity between two figures for classifying various human postures in a video sequence is also presented.

Illumination invariance remains the most researched, yet the most challenging aspect of automatic face recognition. In this paper we investigate the discriminative power of colour-based invariants in the presence of large illumination changes between training and test data, when appearance changes due to cast shadows and non-Lambertian effects are significant. Specifically, there are three main contributions: (i) we employ a more sophisticated photometric model of the camera and show how its parameters can be estimated, (ii) we derive several novel colour-based face invariants, and (iii) on a large database of video sequences we examine and evaluate the largest number of colour-based representations in the literature. Our results suggest that colour invariants do have a substantial discriminative power which may increase the robustness and accuracy of recognition from low resolution images.

We would like to invite you to join this exciting new project as a chapter contributor. Since this is a textbook, a great deal of this chapter entails a survey on the topic under the paradigm of cyber-physical systems, what can be done onboard and remotely, the distributed nature of the system and some exercises on futurology (anticipating trends can shed some light on upcoming designs). IET will bring great visibility to your work. You are welcome to suggest another topic/chapter title if you feel it would be more suitable. Each chapter should be around 20-25 pages each and can be submitted as a Word or Latex File. The IET will send you additional information (formatting, permission form, etc.) with the contributor's agreement once you have agreed to contribute to the book. Visit http:// www.theiet.org/resources/author-hub/books/index.cfm to get all information you need as a contributor to an IET research-level book. Each book is expected to have a total number of 500 printed pages (based on approximately 550 words per page with a 20% allowance for figures and tables). We have included a tentative schedule and list of topics below. If this is something you would consider, please send me the title of your chapter, a short description/abstract of the chapter content, and your full contact details. We will expect original content and new results for this book. You can, of course, reuse published material but the percentage of material reuse for the chapter should be less than 40%. The IET will run a piracy software on the full manuscript to control that you are including original material and will reject chapters who contain a large amount of already-published material so please do take this into consideration.

The human resource (HR) domain contains various types of privacy-sensitive textual data, such as e-mail correspondence and performance appraisal. Doing research on these documents brings several challenges, one of them anonymisation. In this paper, we evaluate the current Dutch text de-identification methods for the HR domain in four steps. First, by updating one of these methods with the latest named entity recognition (NER) models. The result is that the NER model based on the CoNLL 2002 corpus in combination with the BERTje transformer give the best combination for suppressing persons (recall 0.94) and locations (recall 0.82). For suppressing gender, DEDUCE is performing best (recall 0.53). Second NER evaluation is based on both strict de-identification of entities (a person must be suppressed as a person) and third evaluation on a loose sense of de-identification (no matter what how a person is suppressed, as long it is suppressed). In the fourth and last step a new kind of NER dataset is tested for
recognising job titles in tezts.

It is well known that this soiling can reduce the generation efficiency in PV system. In some case according to the literature of loss of energy production in photovoltaic systems can reach up to 50%. In the industry there are various types of cleaning robots, they can substitute the human action, reducing cleaning cost, be used in places where access is difficult, and increasing significantly the gain of the systems. In this paper we present an application of computer vision method for soiling recognition in photovoltaic modules for autonomous cleaning robots. Our method extends classic CV algorithm such Region Growing and the Hough. Additionally, we adopt a pre-processing technique based on Top Hat and Edge detection filters. We have performed a set of experiments to test and validate this method. The article concludes that the developed method can bring more intelligence to photovoltaic cleaning robots.

ABSTRACTThe growth of the telecommunication industry is fast-paced with ground-breaking engineering achievements. Notwithstanding the technological advancement in the industry, it had continued to cope with the phenomenon ofresource constraint in portable mobile telecommunication devices compared to fixed and tethered devices. Portable mobile handheld devices have very low computational, storage and energy carrying capacity occasioned by the needs to satisfy portability, very small form factor, ergonomics, style and trends. Solutions such as cloudlets, cyber foraging, mobile cloud computing (MCC), and more recently but most applicable, multi-access edge computing (MEC) have been proffered with different application methodologies including computational offloading, distributed computing, thin clients, middleware, mobile environment cloning as well as representational state transfer. There is a need to satisfy requirements of new and emerging use cases, especially the deployments of 5...

The hand gesture recognition and pattern recognition are the growing fields of research. Gestures are the motion of the body or physical action form by the user in order to convey some  meaningful information. In this paper, we propose a robust and  efficient method for real-time hand gesture recognition system.
In the suggested method, first, the hand gesture is extracted from  the main image by edge detection and morphological operation and then is sent to feature extraction stage. In feature extraction stage modified contour chain code feature set is extracted. Finally, at classification stage, we employ multiclass support vector machine (SVM) as a classifier. In the result part, the proposed approach is applied to American Sign Language (ASL) database and the accuracy rate obtained 99.40%. Further, we obtained 99.80% accuracy using five-fold cross validation technique on ASL dataset.

Humans recognize thousands of objects, and with relative tolerance to variable retinal inputs. The acquisition of this ability is not fully understood, and it remains an area in which artificial systems have yet to surpass people. We sought to investigate the memory process that supports object recognition. Specifically, we investigated the association of inputs that co-occur over short periods of time. We tested the hypothesis that human perception exploits expectations about object kinematics to limit the scope of association to inputs that are likely to have the same token as a source. In several experiments we exposed participants to images of objects, and we then tested recognition sensitivity. Using motion, we manip- ulated whether successive encounters with an image took place through kinematics that implied the same or a different token as the source of those encounters. Images were injected with noise, or shown at varying orientations, and we included 2 manipulations of motion kinematics. Across all experiments, memory performance was better for images that had been previously encountered with kinematics that implied a single token. A model-based analysis similarly showed greater memory strength when images were shown via kinematics that implied a single token. These results suggest that constraints from physics are built into the mechanisms that support memory about objects. Such constraints—often characterized as ‘Core Knowledge’—are known to support perception and cognition broadly, even in young infants. But they have never been considered as a mechanism for memory with respect to recognition.

International Journal of Computer Vision and machine learning (IJCVML) is an open access, peer-reviewed journal that publishes articles which contribute new results in all areas of the advanced vision computing. The goal of this journal is to bring together researchers and practitioners from academia and industry to focus on understanding advances in vision computing and establishing new collaborations in these areas. Authors are solicited to contribute to the journal by submitting articles that illustrate research results, projects, surveying works and industrial experiences that describe significant advances in the areas of vision computing.

Object recognition is a field in image processing that is widely considered in computer vision and robotics. This paper focuses on recognizing a number of chairs in various shapes and describes them by using chain codes and Fourier descriptor. This paper initially reviews two important steps for the object recognition; namely, feature Extraction and shape description. Then, these approaches are applied to recognize various chairs. It is shown that the chain code and Fourier descriptor are all able to recognize the chairs in general form. However, it is clearly stated that the Fourier descriptor leads to smaller translational, scaling and rotational invariance. In addition, the Fourier descriptor is greatly capable of separating images from their shades while the other two approaches are unable to distinguish real chairs from their respective shapes.

Tomatoes are best known grown fruits in greenhouses. In order to decrease the high labor cost and increase the efficiency, recently many researches have tried to harvest tomatoes automatically. To help the harvesting robot picking the fruit precisely, the main challenges are to identify and localize red mature tomato fruit accurately. In this paper, we presented a robust and effective method for challenging task of segmentation and localization of on-plant tomatoes in the greenhouse, for automatic tomato harvesting purpose. A low-cost RGB-D Kinect V2 camera was used to accomplish the task. Color images and depth images of tomatoes on-plants in the greenhouse were captured in day time using Kinect V2 camera. Along with color and depth images, the corresponding color coordinates, depth coordinates, and camera coordinates were also saved. The proposed method was divided into two stages: fruit segmentation and fruit localization. At the first stage, an improved HSV algorithm and watershed algorithm were used to segment mature fruits from a complex background. An improved HSV algorithm was used to enhance the feature of red mature fruits. After that, Otsu method and morphological operations were used to segment mature fruits from the background and then watershed was implemented to further separate touched fruits. Subsequently, each detected fruit was separately labeled to calculate the pixel coordinates values in color image. At the second stage, a fusion of color and depth information was carried out on the basis of the mapping relationship between color, depth and camera coordinates, to map the obtained fruit pixel coordinates in color image with acquired 3D point cloud data. Then the corresponding point cloud data of each labeled fruit were generated. To remove the noise in the cloud data, a filtration process was performed, and the final point cloud was generated. To verify the proposed method, 18 pairs of images were experimented containing the different numbers of measured fruits with measuring distance, ranging from 0.58 to 1.50 m. The average fruit localization error rate was 1.2 cm for depth distance and 1.9 cm for ground distance. Final results showed that the proposed method has good performance and robustness for segmentation and localization and have good potential to be implemented on the tomato harvesting robot in the greenhouse.

Efficient and accurate object detection has been an important topic in the advancement of computer vision systems. With the advent of machine learning and deep learning techniques, the accuracy for object detection has increased drastically. The project aims to incorporate state-of-the-art technique for object detection with the goal of achieving high accuracy with a real-time performance. In this project, we use a completely machine learning with opencv and deep learning based approach to solve the problem of object detection in an end-to-end fashion. The network is trained on the most challenging publicly available dataset, on which a object detection challenge is conducted annually. The resulting system is fast and accurate, thus aiding those applications which require object detection.

State-of-the-art models for semantic segmentation are based on adaptations of convolutional networks that had originally been designed for image classification. However, dense prediction problems such as semantic segmentation are structurally different from image classification. In this work, we develop a new convolutional network module that is specifically designed for dense prediction. The presented module uses dilated convolutions to systematically aggregate multi-scale contextual information without losing resolution. The architecture is based on the fact that dilated convolutions support exponential expansion of the receptive field without loss of resolution or coverage. We show that the presented context module increases the accuracy of state-of-the-art semantic segmentation systems. In addition, we examine the adaptation of image classification networks to dense prediction and show that simplifying the adapted network can increase accuracy.

In past few decades number of vehicles has increased drastically. So it has become a challenging job to tract them. Even it has become seemingly impossible to identify the car owner in case of violation of any traffic rule or too fast driving. With the explosion in number of vehicles flying on the roads, proper tracking is impossible to be fully managed and monitored by humans; examples are so many like traffic monitoring, tracking stolen cars, managing parking toll, red-light violation enforcement, border and customs checkpoints. Therefore, there is a need to develop Automatic Number Plate Recognition (ANPR) system as a one of the solutions to this problem. Yet it's a very challenging problem, due to the diversity of plate formats, different scales, rotations and non-uniform illumination conditions during image acquisition. The proposed method suggests a histogram based approach which is easily implementable using MATLAB and can duly be verified for it's functionality. This approach has a basic advantage of being simple and hence faster. This paper makes use of various algorithms in each category from distinct edge detection to region of interest extraction. This enhances the performance of the system up to the maximum extent possible with less efforts and use of computational resources.

In this paper we discuss a stand-alone system to allow the deaf or hard-of-hearing people and normal people to communicate with each other easily and fluently. We purpose a model for recognizing Malaysian Sign Language through image processing techniques and converting the visual information into textual information at real-time. In order to implement a real time hand gesture recognition system, video footage will be obtained from digital camera or cellphone camera, then hand position and location will be marked and cropping will be done to isolate them. Then the hand gestures will be recognized through image processing and matched to a prebuilt database of gestures which will be used for textual conversion on the screen. In the final step the normal person will type the text and its equivalent animation of hand gestures will be presented. Through this system, real-time recognition of Malaysian Sign Language and textual representation is done, giving more accurate results at least possible time. It will not only benefit the deaf and dumb people of Malaysia but also could be used in various applications in the technology field. This system provides the flexibility to learn Malaysian Sign Language at personal pace, at anytime and anywhere at home or at workplace.

We present a visual object detector based on a deep convo-lutional neural network that quickly outputs bounding box hypotheses without a separate proposal generation stage [1]. We modify the network for better performance, specialize it for a robotic application involving "bird" and "nest" categories (including the creation of a new dataset for the latter), and extend it to enforce temporal continuity for tracking. The system exhibits very competitive detection accuracy and speed, as well as robust, high-speed tracking on several difficult sequences.

Facial Expression Recognition is very broad domain in the field of Human Computer Interaction. Afterward to improve the recognition rate of within less time it becomes a broader research area. So in this discussion we mainly focus on different Machine Learning and Pattern Recognition techniques to improve in this field. In 21st century different researcher mainly focus in this field. Biometric recognition in the field of computer science has become very intensive interested area now days. Facial Expression Recognition has become very progressive domain because face is easy to interact to be intimate with human. It involves different steps to recognize expression. We discuss these steps below in detail. But it is requisite to detect face, extract feature from detected face then select best feature to recognize and then classify this face. Seven common facial expression include in this area of research. Different data-set and database are freely available like as JAFFEE, Yale and Red Boud etc. Here different techniques involves for detect face, feature extract, select best feature with classification are as PCA(Principal Component Analysis), LBP(Local Binary Pattern), HOG(Histogram Oriented Gradient), SVM(Support Vector Machine), ANN(Artificial Neu-ral Network) and many more are available. All these techniques used to recognize different human expressions in different condition and on various database as mentioned above. We delimit (define) general measures (steps) to recognize human Face Expression. Additionally an effort has been made to represent different techniques in tabular form to reach easily for quick reference.

The thesis concentrates on computational methods pertaining to ancient ostraca - ink on clay inscriptions, written in Hebrew. These texts originate from the biblical kingdoms of Israel and Judah, and dated to the late First Temple period (8th – early 6th centuries BCE). The ostraca are almost the sole remaining epigraphic evidence from the First Temple period and are therefore important for archaeological, historical, linguistic, and religious studies of this era. This “noisy” material offers a fertile ground for the development of various “robust” image analysis, image processing, computer vision and machine learning methods, dealing with the challenging domain of ancient documents’ analysis. The common procedures of modern epigraphers involve manual and labor-intensive steps, facing the risk of unintentionally mixing documentation with interpretation. Therefore, the main goal of this study is establishing a computerized paleographic framework for handling First Temple period epigraphic material. The major research questions, addressed in this thesis are: quality evaluation of manual facsimiles; quality evaluation of ostraca images; automatic binarization of the documents and its subsequent refinement; quality evaluation of binarizations on global and local levels; identification of different writers between inscriptions (two distinct methods are proposed); image segmentation (with improvements over the classical Chan-Vese algorithm); and letters’ shape prior estimation. The developed methods were tested on real-world archaeological and modern data and their results are found to be favorable.

With the availability of enormous amounts of data and the need to computerize visual-based systems, research on object detection has been the focus for the past decade. This need has been accelerated with the increasing computational power and Convolutional Neural Network (CNN) advancements since 2012. With various CNN network architectures available, the You Only Look Once (YOLO) network is popular due to its many reasons, mainly its speed of identification applicable in real-time object identification. Followed by a general introduction of the background and CNN, this paper wishes to review the innovative, yet comparatively simple approach YOLO takes at object detection.

ABSTRACTThe growth of the telecommunication industry is fast-paced with ground-breaking engineering achievements. Notwithstanding the technological advancement in the industry, it had continued to cope with the phenomenon ofresource constraint in portable mobile telecommunication devices compared to fixed and tethered devices. Portable mobile handheld devices have very low computational, storage and energy carrying capacity occasioned by the needs to satisfy portability, very small form factor, ergonomics, style and trends. Solutions such as cloudlets, cyber foraging, mobile cloud computing (MCC), and more recently but most applicable, multi-access edge computing (MEC) have been proffered with different application methodologies including computational offloading, distributed computing, thin clients, middleware, mobile environment cloning as well as representational state transfer. There is a need to satisfy requirements of new and emerging use cases, especially the deployments of 5...

As autonomous vehicles and autonomous racing rise in popularity, so does the need for faster and more accurate detectors. While our naked eyes are able to extract contextual information almost instantly, even from far away, image resolution and computational resources limitations
make detecting smaller objects (that is, objects that occupy a small pixel area in the input image) a truly challenging task for machines and a wide open research field.

This study explores ways in which the popular YOLOv5 object detector can be modified to improve its performance in detecting smaller objects, with a particular focus on its application to autonomous racing. To achieve this, we investigate how replacing certain structural elements of the
model (as well as their connections and other parameters) can affect performance and inference time. In doing so, we propose a series of models at different scales, which we name ‘YOLO-Z’, and which display an improvement of up to 6.9% in mAP when detecting smaller objects at 50% IOU, at a cost of just a 3ms increase in inference time compared to the original YOLOv5.

Our objective is not only to inform future research on the potential of adjusting a popular detector such as YOLOv5 to address specific tasks, but also to provide insights on how specific changes can impact small object detection. Such findings, applied to the wider context of autonomous vehicles, could increase the amount of contextual information available to such systems.

Content based image retrieval (CBIR), also known as query by image content (QBIC) is the application of computer vision to the image retrieval problem, that is, the problem of searching for digital images in large databases. Users of image databases often prefer to
retrieve relevant images by categories. Unfortunately, images are usually indexed by content or low level features like color, texture and shape, which often fail to capture high level concepts well. Again the high level concepts or semantics can vary from user to user. To address these issues, relevance feedback has been extensively used to associate low level image features with high level concepts. Because it is hard to define what will be the most similar images of a query image as it varies from a user to another, taking user judgement on the retrieved images and then refining the search result based on user’s region of interest has become very successful. Many researchers have addressed the issues of Content based Image retrieval with relevance feedback and contributed in this field to a considerable extent. Still there are scope and need of much improvement. Among all existing relevance feedback approaches, query point movement and feature re-weighting have been proven to be suitable for large-scaled image databases with high dimensional image features. In this thesis, we present a query point movement approach using both relevant and irrelevant images to catch the user semantics. The relevance feedback strategy is simple and straightforward. The basic approach adopted here is to move the query point to the direction of user intention. The user only selects the images that he/she thinks relevant and
ignores the irrelevant ones. The feedback provided by the user are categorized into positive and negatives images both of which categories are used in determining the query point for further iterations. During experimentation it is observed that our system can efficiently move towards the user context with minimum possible number of iterations comparing to trivial methods.

Visually impaired people find it difficult to identify objects in front of them while walking down the street, which makes it risky. The smart stick comes as a way to define the world around them.In this paper, we propose a solution, presented in a smart stick with an infrared sensor to detect staircases and a pair of ultrasonic sensors in order to detect any other obstacles in front of the user, within four meters. Furthermore, another sensor is placed at the bottom of the stick to prevent puddles. Speech warning message is triggered upon identification of any obstacle. About 39 million people are permanently blind and about 285 billion are visually impaired. Blind person, who doesn't walk alone, struggles to do his own job. There's always fear of getting injured and getting into an accident. How can we solve those problems? Indeed, this paper has provided a solution which can relieve the blind from this problem. Which is "Blind Stand with voice control to guide blind people" This is essentially a device consisting of Arduino UNO, Ultrasonic Sensor and Voice Module that informs and notifies the blind person of the hazards This will help a blind person to walk comfortably and keep themselves safe from danger.

The paper explains the development of a classification program which is to be integrated to a robot that will autonomously play chess. The problem is to perform a classification on a 12 class data set of chess pieces which works on a real-time video feed. Various approaches have been discussed for the problem; finally Convolutional Neural
Network was chosen to develop the solution.

The paper develops two different approaches to solve the problem. One approach uses a single Convolutional Neural Network for classification. The other divides the problem into two classification task, classifying of the chess piece color and classifying the chess piece based on shape. The models developed from both the approaches are compared based on accuracy and computational time.

Cross validation was used to ensure the classification accuracy results are repeatable. To understand the accuracy in classifying each class, a confusion matrix was generated and the important features were visualized to understand the learnt features. The computation time for all the function calls on the program were measured and the average was calculated. Finally, the robustness of the 1-Step model to the change in object size was experimentally determined.