Research On Object Detection Based On Lightweight Neural Networks

The rapid development of object detection in the past decade benefits from the combination of deep learning,and using convolutional neural networks to extract features has great advantages in speed and accuracy compared with traditional manual extraction.At the same time,the development of target detection can promote the gradual improvement of industrial applications,such as medical image detection,UAV,robot vision,autonomous driving,pedestrian detection,face detection,industrial parts detection and so on,in various fields.But deep learning-based object detection also has various problems in practical applications,first with the efficiency of the network with limited device computing resources--model storage and the speed problems predicted using the model.Secondly,data such as underwater scenes have problems such as noise and target occlusion,these problems need to be solved by continuous optimization of the algorithm and are of important significance in practice.The Anchor-based standard detection algorithm can be divided into single stages and two stages,and the two algorithms have their own advantages.This paper selects the two-stage algorithm as the basic framework,and optimizes the following two problems in this framework:For network efficiency: convolutional neural network efficiency is due to the number of parameters,Especially in the development of computer hardware,The parameter quantity of mainstream convolutional neural network model is huge,This paper improves the network combined with Dense Net network structure and Ghost Net lightweight method,Two dense connection structures are proposed,While exploring a reasonable network width and depth,Is the network parameter quantity calculation quantity and performance to reach the balance.Two-way Dense structures are used to constitute lightweight network LUNet,and serve as the backbone of detection algorithms.On the detection network,the attention fusion module(AFM)of the backbone network is used and tested on the COCO dataset,which improves the network performance.For the data problem: In reality the data is not ideal.There are two problems in the underwater biological datasets used in this paper--data noise and target dense mutual occlusion in the image--greatly compromising the performance of the algorithm.For the first issue,it is proposed to use Retinex as a data enhancement method integrated in the detection algorithm to achieve end-to-end training.After using Retinex,the algorithm m AP value is increased by 0.78.For the latter,the target mutual occlusion is solved by improving the loss function.Results on underwater biological datasets increased 6.72 from the baseline.Experiments on the Pascal VOC dataset show equal improvements in detection on low-dense datasets.