FA-YOLO: Research On Efficient Feature Selection YOLO Improved Algorithm Based On FMDS and AGMF Modules

Yukang Huo,Mingyuan Yao,Qingbin Tian,Tonghao Wang,Ruifeng Wang,Haihua Wang

2024-08-29

Abstract:Over the past few years, the YOLO series of models has emerged as one of the dominant methodologies in the realm of object detection. Many studies have advanced these baseline models by modifying their architectures, enhancing data quality, and developing new loss functions. However, current models still exhibit deficiencies in processing feature maps, such as overlooking the fusion of cross-scale features and a static fusion approach that lacks the capability for dynamic feature adjustment. To address these issues, this paper introduces an efficient Fine-grained Multi-scale Dynamic Selection Module (FMDS Module), which applies a more effective dynamic feature selection and fusion method on fine-grained multi-scale feature maps, significantly enhancing the detection accuracy of small, medium, and large-sized targets in complex environments. Furthermore, this paper proposes an Adaptive Gated Multi-branch Focus Fusion Module (AGMF Module), which utilizes multiple parallel branches to perform complementary fusion of various features captured by the gated unit branch, FMDS Module branch, and TripletAttention branch. This approach further enhances the comprehensiveness, diversity, and integrity of feature fusion. This paper has integrated the FMDS Module, AGMF Module, into Yolov9 to develop a novel object detection model named FA-YOLO. Extensive experimental results show that under identical experimental conditions, FA-YOLO achieves an outstanding 66.1% mean Average Precision (mAP) on the PASCAL VOC 2007 dataset, representing 1.0% improvement over YOLOv9's 65.1%. Additionally, the detection accuracies of FA-YOLO for small, medium, and large targets are 44.1%, 54.6%, and 70.8%, respectively, showing improvements of 2.0%, 3.1%, and 0.9% compared to YOLOv9's 42.1%, 51.5%, and 69.9%.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The paper attempts to address the issue of insufficient accuracy in detecting small, medium, and large targets in complex environments. Specifically, current YOLO series models have some deficiencies when processing feature maps, such as the lack of cross-scale feature fusion and the static fusion methods' lack of dynamic adjustment capabilities, which directly affect the model's effectiveness in complex real-world environments. Additionally, traditional feature fusion structures (such as FPN) cannot transmit information losslessly during cross-layer information transfer, hindering better information fusion.

To tackle these problems, this paper introduces an efficient Fine-grained Multi-scale Dynamic Selection module (FMDS module), which applies more effective dynamic feature selection and fusion methods on fine-grained multi-scale feature maps, significantly improving the detection accuracy of targets of different sizes in complex environments. Meanwhile, the paper also proposes an Adaptive Gated Multi-branch Focused Fusion module (AGMF module), which enhances the comprehensiveness, diversity, and integrity of feature fusion by complementarily fusing various features captured by multiple parallel branches, including the gated unit branch, FMDS module branch, and TripletAttention branch.

With these improvements, the paper develops a new object detection model—FA-YOLO. Experimental results show that under the same experimental conditions, FA-YOLO achieves a mean Average Precision (mAP) of 66.1% on the PASCAL VOC 2007 dataset, which is 1.0% higher than YOLOv9. Specifically, the detection accuracy for small, medium, and large targets increased by 2.0%, 3.1%, and 0.9%, respectively. These results significantly demonstrate the advantages of FA-YOLO in improving detection efficiency and accuracy.