A Dense Face Detection Method Based on Cascaded Multiscale

Abstract

The invention relates to a method for dense face detection based on cascaded multi-scale, which trains detectors of various scale ranges respectively, each object detector is aimed at a specific scale range, and then cascades them to optimize the existing The network structure, this strategy can be carried in any deep model of face detection, has good scalability, and is more suitable for dense small face detection. It can be applied to specific scenarios such as dense crowd monitoring and classroom population statistics, and has strong application value.

Description

technical field

[0001] The invention relates to the fields of deep learning and computer vision, in particular to a method for detecting dense small faces based on cascaded multi-scales. Background technique

[0002] Deep learning has shown great vitality in the fields of image classification and object detection. In the last five years, since AlexNet was proposed, the error rate on the ImageNet dataset has dropped from 15% to 2%, surpassing the human level. On the other hand, in the field of object detection, the best performing detector only achieves 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, and the image size of image classification is often fixed. For convolutional neural networks, a single-scale image is especially suitable for the invariant feature of convolution kernel convolution. Due to the well-trained depth The weight ...

Examples

Embodiment Construction

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] Please refer to figure 1 , the present invention provides a method for dense face detection based on cascaded multi-scale, comprising the following steps:

[0057] Step S1: collect the face data set, and carry out preprocessing to obtain the preprocessed data set;

[0058] Step S2: train a global detector according to the preprocessed data set;

[0059] Step S3: according to the preprocessed data set, construct a local face data set. And train a local detector according to the obtained local face data set;

[0060] Step S5: cascade the global detector and the local detector;

[0061] Step S6: Input the image to be tested into the cascaded global detector and local detector to obtain the global detection result and the local detection result, and combine the global detection result with the local detection result by using the method of non-maximu...

technical field

[0001] The invention relates to the fields of deep learning and computer vision, in particular to a method for detecting dense small faces based on cascaded multi-scales. Background technique

[0002] Deep learning has shown great vitality in the fields of image classification and object detection. In the last five years, since AlexNet was proposed, the error rate on the ImageNet dataset has dropped from 15% to 2%, surpassing the human level. On the other hand, in the field of object detection, the best performing detector only achieves 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, and the image size of image classification is often fixed. For convolutional neural networks, a single-scale image is especially suitable for the invariant feature of convolution kernel convolution. Due to the well-trained depth The weight ...

Examples

Embodiment Construction

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] Please refer to figure 1 , the present invention provides a method for dense face detection based on cascaded multi-scale, comprising the following steps:

[0057] Step S1: collect the face data set, and carry out preprocessing to obtain the preprocessed data set;

[0058] Step S2: train a global detector according to the preprocessed data set;

[0059] Step S3: according to the preprocessed data set, construct a local face data set. And train a local detector according to the obtained local face data set;

[0060] Step S5: cascade the global detector and the local detector;

[0061] Step S6: Input the image to be tested into the cascaded global detector and local detector to obtain the global detection result and the local detection result, and combine the global detection result with the local detection result by using the method of non-maximu...

Images