When the YOLOv5s network extracts features, it treats all model channels equally, which limits the algorithm’s detection effectiveness somewhat. Due to the study’s small and densely dispersed O. hupensis, an effective channel attention mechanism is incorporated into the CSPDarknet53 feature extraction network of YOLOv5s. The enhanced model is dubbed YOLOv5s-ECA-vfnet. The attention mechanism has been demonstrated to be a critical component of enhancing target detection performance and is frequently used in a variety of popular detection algorithms (Cui et al., 2021a; Hu and Yan, 2021). The most prominent feature of Efficient Channel Attention is that it avoids downscaling and cross-channel interactions, while reducing the complexity of the model and enhancing feature representation (Cai et al., 2020; Wang et al., 2020; Zhu et al., 2021). Figure 2 illustrates the structure of the ECA’s attention module.

The ECA module generates channel attention via a fast 1D convolution of size k, the size of which is completely determined by the adaptive channel dimensionality correlation function. After importing feature images with constant dimensionality χ , all channels are globally averaged and pooled. The ECA module will learn features using a one-dimensional convolution that can share weights, and will involve k nearest neighbors per channel to capture cross-channel interactions when learning features. The adaptive k value is determined by the proportional relationship between the masked area of cross-channel information interactions and the channel dimension C, as shown in Eq. 1. K = ( φ ) = | log 2 ( C ) γ + b γ | ood (1) where γ and b are set to 2 and 1 by default, and | * | ood denotes the nearest odd number, and C is the channel dimension.

The general channel attention mechanism module selects a high-level feature map for global pooling and then compresses the two-dimensional features of each channel using dimensionality reduction. After this computation, the entire feature can be considered compressed in order to obtain a complete global perceptual field, but some spatial feature information is lost as the dimensionality is reduced. The relationship between the channel and spatial dimensions cannot be reflected. While more mature CBAM attention mechanisms are based on both space and channel, their computational processes are completely independent, significantly increasing computational time and effort (Wang and Wang, 2021). In comparison, the important significance of Efficient Channel Attention is: Firstly, it is not required dimensionality reduction. all channel and spatial dimension information will be integrated without losing information; secondly, the coverage area and channel dimension through the interaction of cross-channel information. The proportional relationship of C obtains an adaptive k value, which reduces the amount of calculation. The model’s complexity is reduced while the expression of features is increased, increasing the model’s accuracy (Xie et al., 2022b).

The YOLOv5 implements the concept of Auto Learning Bounding Box Anchors, which utilizes the K-Means algorithm to automatically calculate the appropriate anchor box based on the labeled target box (ground truth) (Li et al., 2021a; Cui et al., 2021b), and the anchor box is learned from the training data.

Since the size of the detection target of the custom dataset differs from that of the public dataset, YOLOv5 learns the size of the anchor frame by relearning it automatically. The YOLOv5 generates nine anchor boxes from the COCO dataset: (10, 13), (16, 30), (33, 23), (30, 61), (62, 45), (59, 119), (116, 90), (156, 198) and (373, 326). In this paper, we add three anchoring frame sizes for detecting small targets with inconspicuous boundaries, namely (5, 6), (8, 14), and (15, 11). Anchor boxes were assigned based on the detection layer scale, and the statistical anchor box assignments for detecting O. hupensis with small targets are shown in Table 1.

The YOLOv5 network uses the focal loss function, which has similar loss weights for positive and negative samples superimposed, and the focus is on more difficult to detect targets, which creates some difficulties when training the pinned O. hupensis snails. The number of positive and negative samples is highly unbalanced due to the small size of the pinned O. hupensis snails target. If both positive and negative sample losses are treated equally, the network’s convergence speed will be slowed and the learning information about the target will be lost (Li et al., 2021b). The original loss function concentrates the network’s attention on low-quality samples, reducing the efficiency of updating the target detection model parameters and having a direct effect on subsequent detection results. When the target arrangement is dense, the non-maximum suppression algorithm will eliminate the interference of redundant prediction boxes according to the highest scoring prediction box. At this time, the high score prediction box will affect the subsequent screening results. In this paper, we introduce the VariFocal Loss (Ye, 2021) function for YOLOv5s, as defined in Eq. 2 VFLoss ( p , q ) = { − q ( q ⁡ log ( p ) + ( 1 − q ) log ( 1 − p ) ) q > 0 − α p γ ⁡ log ( 1 − p ) q = 0 (2) In which p is the confidence level of the network output prediction frame, and  q is the target IoU score. For positive samples with higher scores in training, the model increases their share of the loss weights. For negative samples in training, q is 0. The negative sample loss is incorporated with a γ downgraded weight, while the positive sample loss is not introduced in the calculation of γ. The weight of the negative sample loss can be effectively reduced without affecting the positive sample weight. In order to balance q, α downscaling is used and ultimately allows the network to focus its training on the positive samples with high scores.

The deep computing platform used in this paper is comprised of an Intel^® Core(TM) i7-8700K CPU at 3.20 GHz as the central processor, Windows 10 Pro as the operating system, 16 GB DDR4 as the running memory, a GeForce GTX 1070 8 GB as the graphics processor, Python 3.8 with Pytorch 1.9 as the programming language.

The data in this paper were obtained from the study site: Jianghuzhou Island in Zongyang County on the Yangtze River Island in Zongyang County, the north bank of the lower Yangtze River in Anhui Province, China, a dataset of 2000 images of lake and the O. hupensis snails were built under the guidance of blood control professionals from the Tongling Center for Disease Control and Prevention according to the distribution pattern of O. hupensis snails (Zhou, 2019). Some samples of the dataset are shown in Figure 3.

The collected data were filtered and organized, and all O. hupensis image samples in the collected images were annotated using the marker software Make-Sense. The annotated images’ annotation information was saved in the form of.txt files, which included the target object’s category and coordinate information. Figure 4 illustrates a data annotation example.

The average precision (AP) and mean value of average precision mAP are frequently used in target detection to assess the model’s detection effectiveness and performance (Zhao et al., 2021). AP is defined as the area under the Recall and Precision curves, which is equivalent to mAP in this study for a single target. Ratio of area intersections (IoU). The model’s ability to predict the location is determined by comparing the area of the rectangular area predicted by the model to the rectangular area calibrated in the validation set. Precision (P): The ratio of correct targets detected by the model to the total number of targets, which indicates the model’s detection accuracy (Cai et al., 2021b; Shanid and Anitha, 2021). P = TP TP + FP (3) Recall R: Recall is the ratio of targets detected by the model to the total number of targets, and it indicates the model’s ability to identify a wide variety of targets. R = TP TP + FN (4) AP = ∫ 0 1 P ( R ) dR (5) Where TP (True positive) denotes the number of positive samples correctly detected, that is, the prediction frame belongs to the same category as the labeled frame and IoU >0.5. FP (False positive) denotes the number of positive samples incorrectly detected. FN denotes the number of negative samples incorrectly detected. Since precision and recall are affected by confidence level, evaluating model performance solely on the basis of precision and recall would be unscientific and limited. Therefore, in our experiments, we use Average Precision (AP) to evaluate the model’s recognition performance, which is one of the most important indices for evaluating the performance of mainstream target detection algorithms.

The spiked data set is divided randomly into training sets and test sets in a 7:3 ratio. There are 1,400 training and 600 test sets, respectively. The scale of the input image is 640 × 640. The number of training batches is 8. The training momentum is 0.843. The initial learning rate is set to 0.0032. The weight decay is set to 0.00036. The training epochs for theYOLOv5s and the YOLOv5s-ECA-vfnet are both 400 according to the model’s characteristics.