Research On Infrared Scene Target Detection Technology Based On Deep Learning

With the continuous innovation and breakthrough of artificial intelligence technology,computer vision has been gradually integrated into People's Daily life.In recent years,as an important branch of computer vision,target detection has attracted the intense attention of a large number of researchers,and some progress has been made in target detection based on deep learning.However,most of the applications of deep learning are based on the visible light condition,and there are few researches on the infrared scene.Compared with common visible light images,infrared images are very different,with the characteristics of not obvious texture features,fuzzy contour,low imaging contrast,and more noise.These physical characteristics make the target detection in the infrared scene always challenging.In recent years,the convolutional network is used to identify the target,which is less affected by the size change of the target,and the detection result is stable and good.In this paper,considering the use of deep learning algorithm to test the infrared target recognition and deep convolution neural network is used to extract features,through training and learning of a large amount of data to extract various abstract characteristics,and according to the existing algorithm framework did some improvement in the problem,this paper research content mainly includes the following several aspects:In this paper,on the depth of the classical convolution model induction is introduced in detail,based on the current mainstream of deep learning detection algorithm is divided into two broad categories are analyzed,and studied the algorithm framework,based on the database of the most commonly used three kinds of algorithms for testing,the data show that YOLO(you only look once)algorithm both on speed and detection accuracy of certainadvantages,and in between speed and accuracy has achieved a better balance.In practical application,the end to end detection algorithm has reached the need of real-time detection.The collected infrared video was cut out to ensure the diversity of data sources,and the infrared image was processed with histogram to increase the contrast between the target and the background.Based on YOLOv3,the model is trained and tested for the infrared multi-scene data set.The phenomenon of over-fitting and gradient disappearance is improved through the multi-parameter optimization model,which improves the performance of the original model and improves the detection accuracy.The network was improved for the problems exposed in the algorithm.The experimental data showed that the improved algorithm reduced the phenomenon of missed detection and service detection,and improved the detection accuracy as a whole.