Object Detection with Convolution Neural Networks and Custom-SmallerVGGNet

Abstract

AbstractComputer vision has reached significant levels of Object detection. Whether performed via computer vision or deep learning, there are three main results that we to wish obtain, including a list of bounding boxes (or x and y coordinators) for each object in the image, a class label linked with each bounding label, and the probability or confidence score associated with each bounding box and class label. Object detection is becoming even more popular with the breakthrough of deep learning, and especially Convolutional Neural Networks (CNN) which are comprised of two elements, namely convolution layers and pooling layers. With these successes in deep learning, we have more choice on selecting a method and technology to experiment and improve accuracy. There are several datasets such as MNIST, Animal, CIFAR-10, SMILE and ImageNet; these datasets would be a based foundation for researcher to do the experiment. However, there is an approach which can achieve to get a dataset quickly. This paper will introduce two methodologies for object detection and determine the future research directions of researching with object detection in deep dive. First, the traditional object detection pipeline, which will focus on these techniques such as sliding windows, image pyramid and classification. Second, the based network of an object detection framework, this will used certain classification network as a based network in a deep learning object detection such as SmallerVGGNet, Faster R-CNN, SSD or YOLO. Lastly, we experimented using custom SmallerVGGNet and received a positive outcome in so far as greater than 95% of accuracy.KeywordsObject detectionCNNFaster R-CNNSSDYOLOCustom SmallerVGGNet