Research And Application Of Target Detection Algorithm Based On Lightweight Neural Networ

As the basic task of computer vision,object detection has brought great impetus to the development of many downstream tasks and thus has received extensive attention.With the development of deep learning technology,object detection algorithms based on deep neural networks are more inclined to complex frameworks and deeper neural networks,which are very necessary to improve network performance.However,the complex framework and high computational cost also limit the application of many excellent algorithms.This paper aims to think about a reasonable model lightweight method on a concise object detection framework and strive to explore and research it.The first half of this research focuses on the lightweight improvement of keypoint-based object detection algorithms through lightweight module design combined with multi-feature fusion and proposes a multi-feature fusion-based lightweight object detection algorithm(MFFCenter Net).The method utilizes the center point-based Adaptive Gaussian heatmap encoding(AG-Heatmap)strategy proposed in this paper to encode the input to achieve adaptive encoding of the target.The Mix Inverted Residual Module(MIR)proposed in this paper is introduced into the feature extraction network to achieve the effect of lightweight and feature fusion.In addition,a Spatial Pyramid Pooling with Attention Module(SPPA)is introduced to pool,cascade,and filter multi-scale local region features so that the network can learn more comprehensive and effective features.Experiments show that MFF-Center Net improves m AP from 72.56% to 73.71% on the general detection dataset PASCAL VOC,while the number of parameters and computation is significantly reduced.The second half of this research focuses on further lightweight the keypoint-based object detection algorithm by combining knowledge distillation with multi-paradigm network design and transferring it to practical applications.Based on the above,a lightweight object detection algorithm based on heatmap knowledge distillation(HDK-Center Net)is proposed.The method firstly constructs a heatmap-based knowledge distillation framework,which guides the student network to learn more generalized features by mining the latent knowledge in the teacher’s network heatmap.Secondly,combining the Transformer design paradigm and the convolutional neural network design paradigm to build a teacher network with advantages in both local features and global modeling.Finally,a heatmap knowledge distillation loss is designed to dynamically adjust the direction of student network learning by balancing the contribution of the teacher network and the ground truth.Experiments show that HDKCenter Net further improves m AP from 73.71% to 74.23% on the general detection dataset PASCAL VOC while maintaining lightweight,and achieving 73.5% m AP on the vehicle and pedestrian detection dataset L-KITTI.