LA-DeepLab V3+: A Novel Counting Network for Pigs

Chengqi Liu,Longhe Wang,Lin Li,Shuhan Lu,Jie Su

doi:10.3390/agriculture12020284

Chengqi Liu, Longhe Wang + Show 3 more

Open Access

https://doi.org/10.3390/agriculture12020284

Copy DOI

Abstract

Accurate identification and intelligent counting of pig herds can effectively improve the level of fine management of pig farms. A semantic segmentation and counting network was proposed in this study to improve the segmentation accuracy and counting efficiency of pigs in complex image segmentation. In this study, we built our own datasets of pigs under different scenarios, and set three levels of number detection difficulty—namely, lightweight, middleweight, and heavyweight. First, an image segmentation model of a small sample of pigs was established based on the DeepLab V3+ deep learning method to reduce the training cost and obtain initial features. Second, a lightweight attention mechanism was introduced, and attention modules based on rows and columns can accelerate the efficiency of feature calculation and reduce the problem of excessive parameters and feature redundancy caused by network depth. Third, a recursive cascade method was used to optimize the fusion of high- and low-frequency features for mining potential semantic information. Finally, the improved model was integrated to build a graphical platform for the accurate counting of pigs. Compared with FCNNs, U-Net, SegNet, and DenseNet methods, the DeepLab V3+ experimental results show that the values of the comprehensive evaluation indices P, R, AP, F1-score, and MIoU of LA-DeepLab V3+ (single tag) are higher than those of other semantic segmentation models, at 86.04%, 75.06%, 78.67%, 0.8, and 76.31%, respectively. The P, AP, and MIoU values of LA-DeepLab V3+ (multiple tags) are also higher than those of other models, at 88.36%, 76.75%, and 74.62%, respectively. The segmentation accuracy of pig images with simple backgrounds reaches 99%. The pressure test of the counting network can calculate the number of pigs with a maximum of 50, which meets the requirements of free-range breeding in standard piggeries. The model has strong generalization ability in pig herd detection under different scenarios, which can serve as a reference for intelligent pig farm management and animal life research.

Highlights

Introduction iationsGroup free-range breeding will be the mainstream breeding method of pig farms in the future, and the increase in the number of pigs will lead to an increase in manual inspection [1]
The pressure test of the counting network can calculate the number of pigs with a maximum of 50, which meets the requirements of free-range breeding in standard piggeries
The segmentation model based on dilated convolution could enlarge the local receptive field of the original convolution kernel; the proportion of some pig targets in the overall image was small; these two segmentation methods were still imperfect in the performance method

Summary

Introduction

Introduction iationsGroup free-range breeding will be the mainstream breeding method of pig farms in the future, and the increase in the number of pigs will lead to an increase in manual inspection [1]. The achievement of automatic pig identification, trajectory tracking, and quantity statistics by using computer vision technology has become a current research hotspot [2]. In this field, foreground segmentation of pig herd images and separation of adhesive individual images are the basis for achieving automatic inventory of pig numbers [3]. Owing to the complexity of pig images, such as light changes, crowding, stacking, and occlusion, the existing semantic segmentation technology still faces problems, such as missing segmentation and mis-segmentation, which result in inaccurate counting [4,5].

Methods

Results

Discussion

Conclusion