A fast and accurate single-stage target detection method and device

Abstract

The invention discloses a fast and accurate single-stage target detection method and device, abandoning the improvement method of complex bottom layer and high layer network structure, and only optimizing the mainstream structure and adding a lightweight shunt structure (ie shunt structure) to enhance Coordination and unity between multi-layer regression feature maps in high-level networks. The present invention uses a high-level network to perform maximum pooling and asynchronous convolution decomposition operations on the features output by the bottom network. While reducing the dimension of the feature map, it is conducive to the transmission of spatially related information, which can improve the problem of missing spatially related information and improve the quality of features. Diversity and difference. After adding the shunt structure and optimizing the mainstream structure, the average accuracy of target detection results reaches 80.5%, which is 3.3% higher than SSD and 1.9% higher than DSSD321. At the same time, an average processing speed of 30fps is achieved on a 1080ti graphics card.

Description

technical field

[0001] The invention belongs to the technical field of target detection, and in particular relates to a fast and accurate single-stage target detection method and device. Background technique

[0002] As a key technology in the field of computer vision, object detection has always been a challenging research hotspot. The Viola-Jones algorithm proposed by Viola and Jones in 2001 realized real-time face detection for the first time under the condition of limited computing resources. In 2005, the HOG pedestrian detector proposed by Dalal and Triggs expanded the detection field of Viola-Jones algorithm. The deformable part model DMP (Deformable Part based Model) proposed by Felzenszwalb et al. and its follow-up optimization algorithm have won the championship of the VOC Object Detection Challenge (The PASCAL Visual Object Classes Challenge) for three consecutive years, representing the detector based on hand-designed features at that time the highest level. Ho...

Examples

Embodiment Construction

[0063] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0064] The underlying network of the FA-SSD network in the present invention is based on a convolutional neural network (VGG), and the high-level network is a multi-level regression calculation structure. The schematic diagram of the FA-SSD network framework in the present invention is as follows figure 2 shown. figure 2 In , the dotted line box is the mainstream structure of the high-level network, the left side of the dotted line box is the bottom network, and there are channel one, channel two, channel three, channel four and channel five connected in sequence and with the same structure in the dotted line box. figure 2 Among them, circles, triangles, and prisms whose height is greater than width represent the network structure (nn) between convolution operations (conv), pooling operations (pool) and regression feature layers, respectiv...

technical field

[0001] The invention belongs to the technical field of target detection, and in particular relates to a fast and accurate single-stage target detection method and device. Background technique

[0002] As a key technology in the field of computer vision, object detection has always been a challenging research hotspot. The Viola-Jones algorithm proposed by Viola and Jones in 2001 realized real-time face detection for the first time under the condition of limited computing resources. In 2005, the HOG pedestrian detector proposed by Dalal and Triggs expanded the detection field of Viola-Jones algorithm. The deformable part model DMP (Deformable Part based Model) proposed by Felzenszwalb et al. and its follow-up optimization algorithm have won the championship of the VOC Object Detection Challenge (The PASCAL Visual Object Classes Challenge) for three consecutive years, representing the detector based on hand-designed features at that time the highest level. Ho...

Examples

Embodiment Construction

[0063] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0064] The underlying network of the FA-SSD network in the present invention is based on a convolutional neural network (VGG), and the high-level network is a multi-level regression calculation structure. The schematic diagram of the FA-SSD network framework in the present invention is as follows figure 2 shown. figure 2 In , the dotted line box is the mainstream structure of the high-level network, the left side of the dotted line box is the bottom network, and there are channel one, channel two, channel three, channel four and channel five connected in sequence and with the same structure in the dotted line box. figure 2 Among them, circles, triangles, and prisms whose height is greater than width represent the network structure (nn) between convolution operations (conv), pooling operations (pool) and regression feature layers, respectiv...

Images