Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
2394-9007
Vol. V, No. I, February 2018 www.ijrtonline.org
Abstract— General purpose object detection should be fast, Going one step further from here, we would want to not
accurate, and able to recognize a wide variety of objects. Since only find objects inside an image, but find a pixel by pixel
the introduction of neural networks, detection frameworks have mask of each of the detected objects. This problem is referred
become increasingly fast and accurate. However, most detection as Instance or Object segmentation. Iterating over the problem
methods are still constrained to a small set of objects. Use of
of localization plus classification ends up with the need for
traditional approach which is repurposing classifier or localizer
to perform detection applies the model to an image at multiple detecting and classifying multiple objects at the same time.
location and scales. This paper discusses the parameters which Object detection is the problem of finding and classifying a
will enhance the effectiveness of object detection approaches in variable number of objects on an image. The process of object
application areas of object detection. The viewpoint for object detection is done via methods like Viola-Jones Detection,
detection that this paper proposes is parameters like an efficient Feature-based Detection, SVM Classifier with HOG features
dataset, a better feature extractor, localization of bounding boxes and Deep learning Object Detection. Today, the fields of Deep
and dimension adaptability. These features play an important Learning - Artificial Neural networks are being used
role in behavior of any object detection model in terms of ingeniously in Object Detection in real-time such as self
performance, accuracy, and classification process & outcome and
driving cars, assisted living applications, visual search engines
time factor.
and aerial image analysis etc. In this paper, the aim is to
Keywords: Object Detection, Neural Net, Image Recognition, develop a faster deep neural network for real-time video object
Image Classification.
detection by exploring the ideas of knowledge-guided training
I. INTRODUCTION and predicted regions of interest. Specifically, a new
framework is proposed for training neural networks on
Object detection comes under the field of Computer Vision. It classification as well as detection datasets, and different
is the process of finding instances of real-world objects such approach of k-means clustering toward dimension box
as faces, bicycles, and buildings in images or videos. Object clustering, localization in prediction of region proposals and
detection algorithms typically use extracted features and various resolution objects towards making object detection
learning algorithms to recognize instances of an object much more effective, faster and accurate.
category. It all starts from Classification. Classification is the
probably the most well known problem in computer vision. It II. RELATED WORK
consists of classifying an image into one of many different
Object Detection was done classically with Viola-Jones
categories. In recent years classification models have
framework but recent use of deep learning and artificial neural
surpassed human performance and it has been considered
network in detecting and classifying the images have
practically solved. Similar to classification, Localization finds
surpassed the humans as well in Image Net Challenge. In
the location of a single object inside the image. Localization
detection the most used form of neural network is R-CNN.
can be used for lots of useful real-life problems. For example,
Today, R-CNN have been effectively been mastered by Fast
smart cropping (knowing where to crop images based on
R-CNN which has even surpassed Faster R-CNN. And in
where the object is located), or even regular object extraction
recent times, Mask R-CNN has been taking over the detection
for further processing using different techniques. It can be
field in pixel level segmentation.
combined with classification for not only locating the object
Earlier, R-CNN and its successor have used datasets of
but categorizing it into one of many possible categories.
only detection for training their model to perform object
detection. The detection datasets are much limited in size due
Manuscript received on February, 2018.
to unavailability of large, accurately labeled dataset whereas
Shristi Shreyasi, Research Scholar, Department of Computer Science & the classification dataset have millions of images labeled in
Engineering, SRM Institute of Science & Technology, Ramapuram, Chennai, various categories for training models. This becomes barrier in
Tamil Nadu, India.
developing an efficient model for object detection in real time
Prof. S.S Subashka Ramesh, Asst. Professor, Department of Computer videos. Another, main cause for complex and time taking
Science & Engineering, SRM Institute of Science & Technology, outcomes is the poorly predicted bounding boxes in the frame
Ramapuram, Chennai, Tamil Nadu, India.