Research On Lightweight Crested Ibis Detection Algorithm Based On YOLOv5

Crested ibis is a first-class protected animal in China,and the collection of its distribution information is of great significance.Currently,the collection of Crested Ibis distribution information mainly relies on artificial field patrols,resulting in high missed inspection rates and high labor costs.Computer vision technology based on deep learning has made breakthrough progress,and has important research value and broad application prospects in the task of collecting Crested Ibis distribution information.However,high-precision target detection algorithms often have problems such as large model parameters and high computational costs,which are difficult to meet the requirements of lightweight deployment of Crested Ibis detection in the field.This study first designs a lightweight Crested Ibis detection algorithm,and then uses model compression technology to reduce the amount of parameters and computation of the algorithm,thereby reducing the hardware requirements of the algorithm.The main tasks include(1)Aiming at the problems of large parameters and large computational complexity in current detection algorithms,this paper designs a lightweight network model based on YOLOv5.Firstly,combining MBConv Block in Efficient Net network,the original backbone network is reconstructed to significantly reduce network parameters;At the same time,the Stem module is used in the shallow network to improve the feature extraction ability of the shallow network;Then,we improved the convolutional attention module(CBAM),replacing the channel attention with an efficient channel attention module(ECA),avoiding dimensionality reduction operations,effectively extracting information between adjacent channels,and significantly reducing the parameter amount of channel attention.We embedded it in the feature fusion network path aggregation network(PANET),achieving the goal of introducing small parameter amounts and effectively improving network performance,And named it the Efficient Convolutional Attention Module(ECBAM).Compared with YOLOv5 s algorithm,this algorithm achieves a 52.37% reduction in parameter size and a 54.55%reduction in computational complexity.Finally,experiments were conducted on a self built Crested Ibis dataset and a public dataset.The experimental results show that the proposed algorithm and YOLOv5 s algorithm have comparable detection accuracy,and can significantly reduce the amount of network model parameters and computation while maintaining high detection performance.(2)To further reduce the parameter and computational complexity of the algorithm in this paper,a channel pruning algorithm based on the batch standardization layer is used to further compress the above algorithm.Firstly,using L1 regularization based on batch standardization layer to sparsely train the network,change the network weight distribution,and improve the generalization ability of the network model to obtain a batch standardization layer with sparse weights.Subsequently,network model pruning experiments with different pruning rates were conducted.The experimental results showed that,with increasing pruning rates,the amount of parameters and computation of the network model also significantly decreased,but the detection accuracy also gradually decreased,m AP@0.5 : 0.95 The loss is significant,but m AP@0.5 The loss is small,and although the pruning rate is high,the fine-tuned model can still have high detection accuracy.The algorithm achieves a large compression of algorithm parameters and computation at the cost of losing a small amount of detection accuracy.Experimental results verify the effectiveness of the model compression algorithm and related parameter selection used in this article.(3)A crested ibis detection system was built and tested on an embedded device Jetson Nano,including the Jetson Nano,camera head,I/O equipment,and human-computer interaction interface.The experimental results show that the detection frame rate is effectively improved when the pruning rate is 0.5 and 0.8,and the feasibility and effectiveness of the algorithm in practical applications are verified.