Swin-YOLOv3 Improving YOLOv3 network for grain insects detection

Abstract

Abstract Aiming at the problems of low target detection rate and high false alarm rate originating from incomplete edge target, blurred target and overlapping target in grain insect image, one improved YOLOv3, reffered as Swin-YOLOv3, is presented. In the traditional YOLOv3, the receptive field of the ResNet feature extraction network is limited and powerless when facing the smaller grain insect target, and its sensitivity to the local detail features is also poorer. Thus, the Swin Transformer block structure is introduced to increase the receptive field by the shift window operation and enhance the communication ability between network layers, so as to obtain more local feature information; Besides, due to the diversity of grain insects, the difference of their shapes and sizes inevitably causes the conflicts between features at different levels. These conflicts interfere with the gradient calculation in the training process, reducing the effectiveness of the feature pyramid in the traditional YOLOv3 feature fusion module. Adaptive Spatial Feature Fusion(ASFF) has the advantage of adaptive multi-scale feature fusion, thereby, the ASFF is adopted to fuse the features of each layer and adjust them to the same resolution, besides, the optimal fusion can be obtained through training; finally, the downsampled ResNet residual block in the traditional YOLOv3 is replaced by the RepVGG structure, which adapts the convolution layer of the straight tube single path structure with a high degree of parallelism, so that the memory occupancy of the network can be increased obviously and the network training and detection speed is improved accordingly. The experimental results show that, in the public AI insect identification dataset, the mean precision (mAP) of the proposed Swin-YOLOv3 is 89.9%, which is 12.2% and 5.6% higher than the tranditional YOLOv3 and the YOLOv5, respectively; In the self-made dense grain insect dataset, the mAP of the proposed Swin-YOLOv3 is 77.1%, which is 12.4%, 3.8% and 4.6% higher than the traditional YOLOv3, the Swin Transformer and the YOLOv5, respectively; moreover, the proposed Swin-YOLOv3 has a higher detection speed of 36 frames/second.