Target detection method and system based on multi-scale feature map reconstruction and knowledge distillation

Abstract

The invention discloses a target detection method and system based on multi-scale feature map reconstruction and knowledge distillation. The method comprises the steps that firstly, a backbone networkDarkne-53 is utilized to extract features, and deep features are subjected to up-sampling and shallow feature tensor splicing to generate a multi-scale feature map; then, a feature re-calibration strategy is adopted to automatically obtain the weight of each channel in the feature map, useful features are improved according to the weights, useless features are inhibited, and semantic informationof top-layer features and detail information of bottom-layer features are fused through a residual error module; the gamma coefficients of the batch normalization layer in the backbone network are introduced into a pruning target function for training, and a channel where the gamma coefficients lower than a threshold are located is removed from the model according to a pruning threshold; and finally, the trained YOLOv3 reference model is taken as a teacher network, and the pruned model is taken as a student network to perform knowledge distillation. The method and system improve the detectionprecision of objects with different sizes in a large range, reduce the calculation amount of the model, and improve the detection speed of the model.

Description

technical field [0001] The invention provides a target detection method and system based on multi-scale feature map reconstruction and knowledge distillation, belonging to the technical field of target detection in computer vision. Background technique [0002] Image target recognition is a research topic involving computer vision, pattern recognition and artificial intelligence. With the rapid development of hardware technology, embedded smart devices based on deep learning platforms are becoming more and more mature, and more and more detection algorithms are embedded in to smart devices, but the traditional detection method has a large difference in the detection accuracy of targets of different sizes within a certain range, cannot accurately identify targets, and cannot meet daily needs, and the traditional detection algorithm has too many model parameters, and the required computing power is relatively large. Therefore, it is necessary to propose a technology that can n...

AI Technical Summary

Problems solved by technology

Low-resolution, semantically strong features are upsampled and combined with high-resolution, semantically weak features to build a feature pyramid that shares rich semantics at all levels, but it still has a lot of room for improvement, For example, [Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018] manually designed the fusion structure and enhanced feature fusion to improve the detection accuracy a lot, but these algorithms It has not been optimized in combination with YOLOv3 and the actual scene, and there is still a lot of room for improvement in feature map reconstruction

[0005] For the model compression method of target detection, many previous works have been proposed to compress large CNNs or directly learn more effective CNN models for fast reasoning, such as literature [E.L.Denton, W.Zaremba, J.Bruna, Y.LeCun, and R.Fergus.Exploiting linear structure within convolutional networks efficient evaluation.In NIPS,2014.] Applied low-rank approximation, literature [S.Han,J.Pool,J.Tran,and W.Dally.Learning both weights and connections for Efficient neural network.In NIPS, pages 1135–1143,2015.] used weight pruning, etc., but most of these techniques require specially designed software / hardware accelerators to accelerate execution, model compression for target detection on embedded devices There are relatively few methods, and the current compression algorithm compresses YOLOv3 with high time complexity, which cannot well complete the target detection tasks in embedded device application scenarios (such as pedestrian and vehicle target detection in intelligent transportation)

Images