IO-YOLOv5: Improved Pig Detection under Various Illuminations and Heavy Occlusion

Abstract

Accurate detection and counting of live pigs are integral to scientific breeding and production in intelligent agriculture. However, existing pig counting methods are challenged by heavy occlusion and varying illumination conditions. To overcome these challenges, we proposed IO-YOLOv5 (Illumination-Occlusion YOLOv5), an improved network that expands on the YOLOv5 framework with three key contributions. Firstly, we introduced the Simple Attention Receptive Field Block (SARFB) module to expand the receptive field and give greater weight to important features at different levels. The Ghost Spatial Pyramid Pooling Fast Cross Stage Partial Connections (GSPPFC) module was also introduced to enhance model feature reuse and information flow. Secondly, we optimized the loss function by using Varifocal Loss to improve the model’s learning ability on high-quality and challenging samples. Thirdly, we proposed a public dataset consisting of 1270 images and 15,672 pig labels. Experiments demonstrated that IO-YOLOv5 achieved a mean average precision (mAP) of 90.8% and a precision of 86.4%, surpassing the baseline model by 2.2% and 3.7% respectively. By using a model ensemble and test time augmentation, we further improved the mAP to 92.6%, which is a 4% improvement over the baseline model. Extensive experiments showed that IO-YOLOv5 exhibits excellent performance in pig recognition, particularly under heavy occlusion and various illuminations. These results provide a strong foundation for pig recognition in complex breeding environments.

YOLOv5-SA-FC: A Novel Pig Detection and Counting Method Based on Shuffle Attention and Focal Complete Intersection over Union

“…We also found some existing research used in the detection of pigs with YOLOv5 (e.g., Lai [38], Li [39], and Zhou [40]) and compared them with our method. The specific results are shown in Table 6 and were assessed using the mAP@0.5 metric: From Table 6, it can be seen that our YOLOv5-SA performs the best.…”

SAFP-YOLO: Enhanced Object Detection Speed Using Spatial Attention-Based Filter Pruning

“…Attention for Accuracy (Not Report Speed/Model Size) [13] Attention for Accuracy (Not Report Speed/Model Size) [25] Attention for Accuracy (Speed Degradation) [26] Attention for Accuracy (Model Size Increase) [39] Pruning for Speed (Not Report Accuracy) [18] Spatial Attention-based Pruning for Accuracy and Speed Proposed Method…”

“…In general, there exists an accuracy/speed trade-off, and therefore, introducing attention mechanisms to improve accuracy increases the resource requirements (time and memory), while pruning for speed sacrifices accuracy. For example, in [26], applying attention improved accuracy from 90.4% to 90.8%; however, the speed decreased slightly from 83FPS to 82FPS. Similarly, in [39], the application of attention improved accuracy from 95.2% to 97.2%; however, it was reported that the model size increased slightly from 13.7MB to 14.4MB.…”