A Dynamic Multi-Label Image Classification using Recurrent Neural Network

Abstract

Various developmental standards have been devised for neural network applications. Tragically these methods are not able to scale up to modern standards set by neural system because of entangled design and enormous number of loads. In this paper we propose a new method for developing neural system according to the evolutionary standards using Recurrent Neural Network, this technique is Image Classification, which is used for identifying various features of an image and classifying images according to its visual content. The motive behind this project is solving day to day problems concerning images. For example, obstacle detection to prevent accidents, handwriting recognition. These varied forms of problems are redefined and given a peculiar structure to ease classification and recognition. RNN studies a combined image-name implant to characterize semantic reliance and image pertinence. The RNN perform the task recursively for every element of a sequence, with the output being dependent on the past calculations. Thus, the system learns from the previously trained model making it easier and faster for computations, it deals with noisy inputs efficiently and does not face the problem of overfitting. The experimental results demonstrate the momentous predominance of proposed algorithm in terms of accuracy, loss rate, number of parameters and hence provide an efficient generalization capacity.