Improving YOLOv8 for Microbial Object Detection in Microscope Images

To address the issues such as large scale variation, edges blurring, high density and overlap of microbial objects in the high-resolution microorganism images, this paper proposes an enhanced object detection model IM-YOLOv8 tailored for microscope images. Firstly, the super token attention is introduced into the Bottleneck module of the CSPLayer for feature extraction. This attention module can capture the global relationship based on super pixels in the early stage of the neural network, and improve the feature discrimination ability, so as to deal with the problem of ambiguity in object boundaries. Secondly, a progressive adaptive feature fusion module is designed to gradually fuse the low-level features to generate more discriminative multi-scale features. Finally, these two modules are assembled into the YOLOv8 model to obtain the enhanced object detection model for microscope images. The experimental results on the AGAR-HR dataset demonstrate that our method is 1.6% and 3.2% higher than YOLOv8 in terms of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{mAP}_{50}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{mAP}_{50:95}$</tex> respectively. In addition, the proposed method possesses a more compact structure, and the model parameters are 16.4 M less than YOLOv8, which proves the superior performance of this method.