Faster R-CNN Network Based on Multi Feature Fusion for Efficient Face Detection

Abstract

The putting into effect of deep learning gave all attention on deep convolutionary neural networks has gained great condition of having general approval in face discovery in near in time years. One of the still in the same way open tasks however is to make out small-scale points. The distance down of the convolutionary network can cause a quick becoming-smaller of the sent out point map for small faces, and most scale unchanging views can hardly grip less than 15×15 bit of picture faces. There are few types of Haar-like rectangle points, which send in to the hard question that the training time for the classifier is too long because of, in relation to the greatly sized number of point amounts needed to put in order the small faces. In order to get answer to this hard question, we offer the MB-LBP (Multi-scale Local based on good example) features and joined rotation-invariant LBP (nearby based on good example) features based on the quicker R-CNN classifier called convolutional neural network (put) in middle in the gave greater value to quicker field, range (EFR-CNN). Despite the shortcomings of MB-LBP point and rotation-invariant LBP point on face edge knowledge, the canny operator-based edge azimuth field purpose, use is grouped together with the above features to make statement of the sense of words of the face news given. In place, based on the grouped together point group, a given to overmuch pleasure R-CNN network of parallel form is suggested. The proposals are put on one side to three being like (in some way) given to overmuch pleasure R-CNN networks according to the different rates on a hundred that they cover on the pictures. The three networks are separated by the measures of the map and (make, become, be) different in the weight of the concatenation 8 of purpose, use maps from one another. The offered EFR-CNN system gets done giving undertaking operation on common points of comparison, including FDDB, AFW, PASCAL faces, And wider face, made a comparison of with state-of – the-art face discovery methods such as Unit Box, hyperface Fastcnn.