Improved YOLOv7-tiny nematode detection model for edge devices.

Nematodes are widely used as model organisms in biological researches. To address the challenges during nematode activity screening stage, such as the small size of individual nematode target, easy to be obscured, and the poor lightweight performance and difficult to deploy on edge devices of existing nematode detection models, we proposed an improved YOLOv7-tiny nematode detection model tailored for edge devices. The MobileOne network was employed as the backbone network to boost the model's computational efficiency. The Generalized Feature Pyramid Network(GFPN) was incorporated to refine the Neck layer to enable adaptive fusion of "skip-layer" and "cross-scale" approaches, thereby enriching the representation of image features. Additionally, a dual-layer routing attention mechanism(BRA) was introduced into the Neck layer to enhance the feature extraction capability for obscured targets. The fourth detection head was added into the Head layer to enhance the detection capability for small targets. The INT8 quantization processing was adopted for the model using the perceptual quantization method, with an asymmetric quantization strategy applied to the activation values to further reduce computational load and achieve model lightweighting. The improved model was deployed and tested on the edge device Jetson Nano. The experimental results indicated that compared to the original model, the improved model showed an increase in mean average precision(mAP@0.5) by 2.7 percentage points, a reduction in computational demand(GFLOPs) by 67.71%, and an increase in detection frame rate(FPS) by 23.01%. These results demonstrated that the accuracy of the improved model was significantly enhanced, and it could be enabled rapid and precise detection of nematode targets on edge devices.