A deep learning-based method for silkworm egg counting

Abstract

The counting of silkworm eggs is an essential task in the selection and breeding of new silkworm species, as well as in silkworm egg production. Currently, this task mainly relies on manual counting, which poses many challenges such as high workload, low efficiency, and being error-prone. To alleviate these problems, this study proposes a deep learning-based method for silkworm egg counting. Images of silkworm eggs were captured from actual environments and annotated using a labeling tool, resulting in more than 300,000 labeled eggs. A counting network based on the You Only Look Once (YOLOv8n) object detection network is proposed, in which an efficient multi-scale attention (EMA) module is embedded in the extraction block of the original network to enhance feature representation capability and suppress interference. To further improve counting performance, a space-to-depth convolution (SPD-Conv) block is introduced to replace the down-sample layer implemented by convolutional layers with a stride of 2. The proposed network is termed YOLO for silkworm egg counting (YOLO-SEC). Experimental results demonstrate that our YOLO-SEC achieves a recall of 99.50 %, a precision of 98.29 %, an F1-score of 0.99, and an AP of 99.31 % for silkworm eggs on the test set. Meanwhile, YOLO-SEC shows significant performance advantages over the original YOLOv8 (n ∼ x), state-of-the-art networks (YOLOv7-tiny, YOLOv9s, YOLOv10n and YOLOv11n), and related networks (YOLOv8-QR, EDGC-YOLO, Faster-YOLO-AP, YOLOv8-ECFS). This research provides technical support for the breeding and egg production process of silkworms.