Target detection performance optimization method

Abstract

The invention discloses a target detection performance optimization method. The method comprises the steps that in the training process of a detection model, metric learning is used for adjusting distribution of samples in a characteristic space to generate characteristics with higher discrimination; a deep neural network corresponding to metric learning is in iteration training, and a candidate box used in each iteration is determined through united overlapping of IoU information and has the positional relation that identical target object distances meet a certain constraint condition and different target distances meet a certain constraint condition; whether the characteristics of a candidate box target generated in each turn of iteration training meet a similarity constraint condition or not is checked; if yes, the detection model does not generate loss in this iteration, and output errors corresponding to all layers in a reverse propagation network are not needed; and during testing, a picture to be detected and a candidate box set of the picture are input into the detection model obtained after training to obtain target object coordinate and category information output by the detection model. Through the method, detection capability can be improved, and detection performance can be optimized.

Description

technical field [0001] The invention relates to target detection technology, in particular to a method for optimizing target detection performance. Background technique [0002] Object detection has always been an important research topic in the field of computer vision, and object detection is also the basis of object recognition, tracking, and action recognition. Nowadays, with the successful application of deep neural networks in the field of computer vision, people have invested more research in the field of object detection, such as face detection, pedestrian detection, vehicle detection and so on. [0003] For target detection, the existing mainstream detection frameworks adopt the Object Proposal strategy; first, a series of potential candidate boxes are generated in the picture, and the areas marked by the candidate boxes are potential objects that have nothing to do with the category; secondly, use The detection algorithm extracts the corresponding visual features ...

AI Technical Summary

Problems solved by technology

[0004] However, in practice, due to the limited accuracy of the candidate box generation algorithm, the generated candidate boxes often cannot cover the objects in the picture well, and many candidate boxes only cover part of the object or cover the background with a very similar appearance, which leads to classification. The misjudgment of the classifier may also be that the candidate box includes a part of the background and a part of the target, which leads to the misjudgment of the classifier

Images