Research On Image Detection Algorithm Based On Improved Faster RCNN

In the information age,images can carry more information than text,and deep learning techniques designed for images are becoming increasingly mature.Deep learning techniques applied to images are becoming increasingly prevalent in everyday life,especially in industrial production.More and more industrial production lines are gradually replacing manual work with image detection technology to achieve higher production efficiency and higher detection accuracy.In the past,researchers have conducted a lot of research on the technology of object detection through deep learning.Overall,object detection mainly targets two issues,one is the extraction of image features,obtaining the feature information of the foreground in the image.The method and quality of feature acquisition directly affect the performance of detection;the other is the magnitude of the model,which directly affects the speed of model detection.In response to the above problems,this paper improves the model based on Faster RCNN to obtain a directed Faster RCNN object detection model.The main work of this paper is as follows:In response to the difficulty in obtaining data in industrial production and the disadvantage of incomplete data acquisition,a new target detection data set is proposed.The main part of the data set is a simulated data set of building blocks,obtained by rendering using 3D simulation software,including multiclass simulated data of building blocks from different perspectives.The data can serve as pre-training for Sim2 Real,reducing the training cost of real scenarios.In response to the issue of Faster RCNN’s low efficiency in recognizing similar objects,and to avoid the disadvantage of easy confusion between different similar objects,this paper improves the network architecture.It uses the high-level,mid-level,and low-level semantic information of the image,combines features of different levels to detect objects,and replaces the network structure with an improved dual-channel complementary MAE to achieve higher efficiency and accuracy.The Oriented RPN structure is improved to better suit the building block detection data set.After experimental comparison,an average detection accuracy of 82.4 was finally achieved.Industrial object detection has always been a core issue in industrial production.This paper generates a simulated data set by simulating a real scene,and trains the improved object detection model on this data set.It then fine-tunes on real data that are difficult to obtain,obtaining a target detection algorithm model suitable for real scenarios.Compared with the traditional Faster RCNN method,we have achieved a significant improvement in detection performance.