Gripped target detection algorithm based on edge extraction and cavity convolution

Abstract

The invention discloses a captured target detection algorithm based on edge extraction and cavity convolution, and the algorithm comprises an Edge feature module, a Darknet-53 feature extraction module, a feature pyramid structure, and a Dilation-ASFF network, and the specific steps are as follows: step 1, image collection: collecting an RGB image of a target, and extracting the edge information of the image by using the Edge feature module to form the feature information of four channels; step 2, feature extraction: inputting the collected feature information into a Darknet-53 feature extraction module, performing feature extraction, improving a loss function into a CIoU loss function, and constructing a feature pyramid structure; and step 3, feature fusion: fusing feature information of different scales in the feature pyramid structure through a Dilatation-ASFF network combined with cavity convolution. According to the method, the detection precision and the detection speed of the network are effectively improved, and the sensing capability of the network to multi-scale information and the convergence capability of the network are improved.

Description

technical field [0001] The invention relates to the field of grabbing target detection algorithms, in particular to a grabbing target detection algorithm based on edge extraction and hole convolution. Background technique [0002] As a big manufacturing country in China, the secondary industry occupies an important position in my country's economic structure. In recent years, with the industrial transformation and upgrading of the domestic industrial manufacturing industry, industrial robots have become more and more popular and applied in industrial production. The grasping task of industrial robots mainly includes three important steps: grasping detection, trajectory planning, and control execution. The main task of grasping detection is to use computer vision technology to locate the grasping point and identify the target to be grasped. With the development of deep learning, the target detection algorithm can be roughly divided into two categories: Two-stage target detec...

AI Technical Summary

Problems solved by technology

[0003] The One-stage target detection algorithm does not need to generate candidate frames, and can directly generate the category and location information of the target. Compared with the Two-stage target detection algorithm, it has a huge advantage in speed: Joseph Redmon proposed the first one-stage target detection algorithm in 2015 YOLOv1 network, compared with the two-stage network at the time, the detection speed is extremely fast, but the accuracy is relatively low; the subsequent SSD network has greatly improved the detection accuracy of multi-scale targets; in 2016, the YOLO network launched The second generation, compared with YOLOv1, has a certain improvement in accuracy and speed, but lacks the capture of multi-scale information; in 2018, the YOLOv3 network proposed to replace the feature extraction network with DarkNet53, and replace Softmax with Logistic for object classification , and borrowed from the FPN idea to build a feature pyramid structure, its accuracy is slightly higher than that of the SSD network, but its speed is more than twice that

Images