Today, Environmental Microorganisms (EMs) are inseparable from our lives (Li et al., 2016). Some EMs are conducive to the development of ecology and promote the progress of human civilization (Zhang et al., 2022b). However, some EMs hinder ecological balance and even cause urban water pollution to affect human health (Anand et al., 2021). For example, Oscillatoria is a common EM, which can be observed in various freshwater environments and can thrive in various environments. When it reproduces vigorously, it will produce unpleasant odors, cause water pollution, consume oxygen in the water, and cause fish and shrimp to die of hypoxia (Lu et al., 2021). In addition, Scenedesmus is also a freshwater planktonic algae microorganism, which is usually composed of four to eight cells. Scenedesmus has strong resistance to organic pollutants, and plays a vital role in water self-purification and sewage purification (Kashyap et al., 2021). The images of the EMs proposed above are shown in Figure 1. But researchers using traditional methods of identifying and analyzing microorganisms will consume a lot of manpower and material resources (Ji et al., 2021). Computer vision analysis method is of great significance, it can help researchers to analyze EMs with higher precision and more comprehensive indicators (Kulwa et al., 2022).

Automated microscopes and intelligent microscopes are currently a major trend in development, and everyone is working to develop faster, more accurate and objective hardware and software (Abdulhay et al., 2018). Compared with manual identification and observation methods, computer-aided detection methods are more objective, accurate and convenient. With the rapid development of computer vision and deep learning technologies, computer-aided image analysis has been widely used in many research areas, including histopathology image analysis (Chen et al., 2022a,b; Hu et al., 2022a,b; Li et al., 2022), cytopathology image analysis (Rahaman et al., 2020a, 2021; Liu et al., 2022a,b), object detection (Chen A. et al., 2022; Ma et al., 2022; Zou et al., 2022), microorganism classification (Yang et al., 2022; Zhang et al., 2022a; Zhao et al., 2022a), microorganism segmentation (Zhang et al., 2020, 2021a; Kulwa et al., 2023), and microorganism counting (Zhang et al., 2021b, 2022c,d). In addition, with the advancement of computer hardware and the rapid development of computer-aided detection methods. In addition, with the advancement of computer hardware and the rapid development of computer-aided detection methods, the results obtained by computer-aided detection methods in EM testing have been improving. EMs play a very important role in the whole ecosystem. Because of their small size, invisibility to the naked eye, and unknown nature, studying EMs has always been a challenge for humans (Ma et al., 2022). Generally there are four traditional methods of detecting EMs. The first is physical method, which has a high degree of accuracy, but uses very expensive equipment and a time-consuming analytical process (Yamaguchi et al., 2015). The second method is the chemical method, which has a high identification capacity, but is often affected by environmental contamination from chemical agents. The third molecular biological method, by detecting the genes of microorganisms and analyzing them. The accuracy of this method is very high, but at the same time it consumes a lot of consumes a lot of human and material resources (Kosov et al., 2018). The fourth is morphological method, which requires a skilled researcher to observe the shape of EM under a microscope, which is inefficient and time consuming (Li et al., 2019). These four traditional methods of analyzing EM have their own advantages and disadvantages. As deep learning has been widely used in machine vision analysis in recent years (Shen et al., 2017), it can compensate the shortcomings of the four traditional methods while retaining accuracy (Kulwa et al., 2023). However, there are few public EM databases available to researchers today. This hinders the analysis of EM. In recent years with the advent of the EMDS series of datasets, research on artificial intelligence on EM has been carried out. With the update of EDMS-1 to EMDS-6, environmental microbial images are widely used for classification, segmentation, and retrieval (Zhao et al., 2022b). However, there is still a gap for multi-objective EM detection, and EMDS-7 fills this gap as a multi-objective EM dataset. the EMDS details are listed in Table 3.

Image analysis is the combination of mathematical models and image processing techniques to analyze and extract certain intelligence information (Song et al., 2022). Image processing refers to the use of computers to analyze images. Common image processing includes image denoising, image segmentation, and feature extraction (Liu et al., 2022d). Image noise appears in the process of acquiring and transmitting EMs images (Gonzalez and Woods, 2002). Image denoising can reduce the noise of the EM image while preserving the details of the image (Rahaman et al., 2020b). Besides, in the process of deep learning based EMs image analyzation, we can extract the features of EM images, then send them to the deep learning network model for training, and match them with known data to classify, retrieve and detect EMs (Liu et al., 2022c). In addition, EM images can also be applied in the field of image segmentation to separate microorganisms from the complex background of the image (Pal and Pal, 1993). Meanwhile, EM images can be used in the field of EM object detection. First, we can frame and mark the known EM objects in the original image, and then transfer the image to the object detection model for feature extraction and network training (Zhang et al., 2021b). Finally, the trained model can be applied for object detection of EMs.

EMs are one of the important indicators for investigating the environment, so it is very essential to collect EM data and information (Kosov et al., 2018). Environmental Microorganism Image Dataset SeventhVersion (EMDS-7) are all taken from urban areas, which can be used to monitor the pollution of the urban water environment. Furthermore, due to the constant changes in conditions such as temperature and humidity, EMs are very sensitive to these conditions, so the biomass of EMs are easily affected (Rodriguez et al., 2020). It is difficult to collect enough EM images. Currently, there are some EM data sets, but many of them are not open source. EMDS-7 is provided to researchers as an open source data set. In addition, we prepare high-quality corresponding object label files of EMDS-7 for algorithms and model evaluation. The label file of EMs can be directly used in multiple object detection and analysis. EMDS-7 has a variety of EM images, which provides sufficient data support for EMs object detection and achieves satisfactory detection results. Researchers can apply many artificial intelligence methods instead of traditional analysis methods to analyze microorganisms in EMDS-7.

The main contributions of this paper are as follows.

The object detection model of the YOLO series is a one-stage detection network, which can locate and classify the objects at the same time. The advantage is that the training speed is fast with less time-consuming. One of the most types is YOLOv3. Joseph Redmon and others used the new basic network darknet-53 in the backbone of YOLOv3 for physical sign extraction (Redmon et al., 2016). It contains 53 convolutional layers and introduces a residual structure so that the network can reach a deep level while avoiding the problem of gradient disappearance. In addition, darknet-53 removes the pooling layer and uses a convolutional layer with a step size of 2 to reduce the dimensionality of the feature map, which can maintain the information transmission better. And YOLOv3 also has excellent structures such as anchor and FPN (Lin et al., 2017a).

YOLOv4 is an improved version based on YOLOv3, which adds CSP and PAN structures (Bochkovskiy et al., 2020). The backbone network of YOLOv3 is modified to CSPDarkne53, and add an spp (spatial pyramid pooling) idea behind the backbone network to expand the receptive field, using 1 × 1, 5 × 5, 9 × 9, 13 × 13 as the largest Pooling, multi-scale fusion, and improve the accuracy of the model. At the same time, in the neck network of YOLOv4, there are Feature Pyramid Network (FPN) (Lin et al., 2017a), Path Aggregation Network (PAN), BiFPN, and NAS-FPN, which can collect different feature maps more effectively.

SSD is another striking object detection network after YOLO. SSD has two major advantages (Liu et al., 2016). First, SSD extracts feature maps of different scales for detection. Large-scale feature maps can be used to detect small objects, while small-scale feature maps can be used to detect large objects. Second, SSD uses different Prior boxes (Prior boxes, Default boxes, Anchors) for scale and aspect ratio. It follows the method of direct regression box and classification probability in YOLO, and uses anchors to improve recognition accuracy referring to Faster R-CNN. By combining these two networks, SSD balances the advantages and disadvantages of Faster R-CNN and YOLO.

The RetinaNet object detection model is also a one-stage object detection network. RetinaNet essentially consists of a backbone network (BackBone) and two subnets (SubNet) (Lin et al., 2017b). The backbone network is responsible for calculating the convolution feature map on the entire input image, which is composed of the ResNet residual network and the FPN feature pyramid network. The two sub-networks use the features extracted from the backbone network to achieve their respective functions (Lin et al., 2017a). The first sub-network completes the classification task; the second sub-network completes the bounding box regression task.

Faster RCNN generates candidate frames based on the Anchor mechanism by adding a region proposal networks (RPN), and finally integrates feature extraction, candidate frame selection, frame regression, and classification into one network, thereby the detection accuracy and efficiency can be effectively improved (Ren et al., 2015). Faster RCNN performs classification and detection of foreground and background in the RPN network structure, optimizes the complexity of picking samples, makes positive and negative samples become more balanced, and then focuses on some parameters for classification training. For the first stage of object detection, it has to do both localization and classification, and there is no clear division of labor which part is dedicated to classification and which part is dedicated to regression of prediction frames, so that the learning difficulty increases for each parameter. Therefore, the classification difficulty of the second stage training will be much easier than the first stage target detection to do mixed classification and prediction frame regression directly.