Cascaded multi-scale-based dense face detection method

Abstract

The invention relates to a cascaded multi-scale-based dense face detection method. According to the method, the detectors with various scale ranges are trained respectively, each object detector is cascaded according to a specific scale range to optimize an existing network structure, and the strategy can be carried in a depth model of face detection, has good expansibility, and is more suitable for dense small face detection. The method can be applied to the specific scenes, such as intensive crowd monitoring, classroom people counting, etc., and has very strong application value.

Description

technical field [0001] The invention relates to the fields of deep learning and computer vision, in particular to a cascaded multi-scale dense small face detection method. Background technique [0002] Deep learning has shown great vitality in the fields of image classification and object detection. In the past five years, since AlexNet was proposed, the error rate in the ImageNet dataset has dropped from 15% to 2%, which has surpassed the human level. On the other hand, in the field of object detection, the best-performing detector only reaches 60% mAp in the COCO dataset. Why is object detection relatively difficult for image classification? [0003] This is because object detection is different from image classification tasks. The image size of image classification is often fixed. For convolutional neural networks, single-scale images are especially suitable for the invariant characteristics of convolution kernel convolution. Due to the well-trained depth The weight of ...

AI Technical Summary

Problems solved by technology

This is a perfect attempt of deep learning in the field of image classification, and it exceeds the human level, but unlike the classification task, object detection not only needs to give the calculated object category, but also gives the position information of the object in the image , which undoubtedly increases the difficulty of object detection, and, in an image, there will be objects of different scales, some objects occupy only a few pixels, which further increases the difficulty of object detection

Since there are various difficulties in object detection, and the scale range of objects with the same label may be very large, this is a huge challenge for the scale invariance of convolutional neural networks. In datasets with large scale range changes, a Detectors must accommodate objects of various scales

In addition to the problem of large scale changes, when the density of objects in the image is too dense, there will be missed detection, so this is also one of the problems to be solved

Images