Research On Target Recognition And Location Based On Prior Knowledge Using Deep Learning

In recent years,with the development and advancement of science and technology,artificial intelligence is gradually entering people's lives.Since human beings mainly acquire external information through vision,the development of artificial intelligence is particularly important in visual direction,while target detection is the focus of the field of visual.The face gates and license plate recognition systems that can be seen everywhere are inseparable from the high-precision target detection algorithm.With further research on detection algorithms,the accuracy and recall rate of detection are constantly improving.However,the general target detection method still cannot meet the accuracy requirements in many fields.In order to solve this problem,this paper proposes a target detection and localization method that combines prior knowledge.Under the condition of combining known patterns,reasonable space and other prior knowledge,the accuracy and recall rate of target detection in specific scenarios are improved.The research content is as follows:First,the classification network ResNet and the target detection method Faster-Rcnn are introduced.The structure principle of Resnet network is explained.Then the process of target detection by Faster-Rcnn algorithm is described in detail from Roi-pooling layer,area suggestion network and network loss function.In this paper,the Roi-Align layer is used instead of the Roi-pooling layer,which illustrates the improvement of the positioning accuracy of Roialign and introduces the principle of the FPN network.Secondly,the evaluation indicators of the target detection method are introduced.The analysis finds that the success rate of some target classifications is naturally better than other targets,and it is easy to distinguish from the background.In order to improve the recognition success rate of objects in complex environments and determine the position and posture of objects,this paper introduces patterns as artificial markers in the environment,selects the indoor art decoration patterns,and designs the indexable evaluation indicators of the patterns.A subset of patterns that are confused and easily distinguishable from the background.Thirdly,the principle of using binocular camera for target positioning is introduced.Two patterns of known distance are pasted on a large target,and the binocular parallax is calculated by the method of Faster-Rcnn and template matching,and the target position and posture are settled.With the positional attitude of the large object as the prior knowledge,the space constraint condition of the three-dimensional space is designed,the target detection probability threshold is lowered,and all the detection frames appearing at a reasonable position are found.Finally,the ResNet-S network with HOG features and convolution features is designed.The SENet network is used to calculate the different weights of different feature maps.The HOG features and convolution features are combined to design a feature fusion classifier with adaptive weights.Multiple sub-classifiers are used instead of multi-classifiers,so that the classifier can accurately detect similar objects with small differences and achieve accurate classification.