Publications

Abstract : Object detection has a prominent role in image recognition and identification. Emerging use of neural networks approaches toward image processing, classification and detection for increasing amount of complex datasets. With the collection of large amounts of data, faster and more efficient GPUs and better algorithms, computers can be trained conveniently to detect and classify multiple objects within an image with high accuracy. Single-shot detector-MobileNet (SSD) is predominantly used as it is a gateway to other tasks/problems such as delineating the object boundaries, classifying/categorizing the object, identifying sub-objects, tracking and estimating object's parameters and reconstructing the object. This research demonstrates an approach to train convolutional neural network (CNN) based on multiclass as well as single-class object detection classifiers and then utilize the model to an Android device. SSD achieves a good balance between speed and certainty. SSD runs a convolution network on the image which is fed into the system only once and produces a feature map. SSD on MobileNet has the highest mAP among the models targeted for real-time processing. This algorithm includes SSD architecture and MobileNets for faster process and greater detection ratio.