Research And Application Of Microalgae Lightweight Detection Algorithm Based On Improved YOLO V7

Microalgae are a group of low-level aquatic microorganisms that are autotrophic through photosynthesis.Microalgae play an important role in aquaculture,water quality monitoring,food processing,biofuel and environmental protection.Over the years,microalgae detection has been getting more and more attention,and it is great of importance to perform real-time,rapid and accurate intelligent detection of microalgae.Microalgae are extremely tiny in size and cannot be directly observed by the human eye,so it is necessary to use a microscope to clearly see their morphology.However,the current data set of microscopic images of microalgae is very scarce.Some of the microalgae detection methods based on deep learning have high detection accuracy,but they are pretty complex and can identify fewer microalgae species;while the detection accuracy of microalgae detection methods that can identify more species is low and these methods spend more time,which cannot meet the demand of microalgae detection in real scenarios.In this paper,in order to solve the problems caused by the sparse samples of microalgae datasets and to equilibrize the relationship among detection accurateness,computational complicatedness and detection velocity of microalgae detection algorithms,a lightweight real-time detection method YOLO v7-MA was proposed for the specific task of microalgae detection,combined with microalgae microscopic image features and based on the YOLO v7 model.And based on the YOLO v7-MA model,an intelligent online detection platform for microalgae was developed to provide reference for relevant inspectors and further reduce the workload of inspectors.The concrete study of this thesis is as follows.(1)Construction and data enhancement of microalgae dataset.To address the problem of insufficient sample size of the current microalgae dataset,this paper further selected data from the EMDS-7 dataset,which originally contained 2365 microbial microscopic images and 41 microbial categories,to form a microalgae dataset with1512 microscopic images of 14 microalgae species.On this basis,data enhancement was then applied to the microalgae dataset to enlarge the sample scale of the dataset and improve the ability of generalization and robustness of the microalgae detection model.There were two main ways of data augmentation,the offline augmentation and the online augmentation.Offline enhancement mainly expanded the training set by random combinations of Gaussian blur,horizontal flip,vertical flip,non-equal scaling,random panning,perspective transformation and random cropping;online enhancement mainly increased the diversity of data samples during model training by both Mosaic and Mixup.After data augmentation,a total of 15,935 microalgal microscopic images were obtained.The comparison experiments based on YOLO v7 showed that the m AP of the model was 97.61%after using offline enhancement and online enhancement,which was a significant improvement of 41.23%compared with the model without data enhancement.(2)Microalgae detection model lightweight improvement and detection performance enhancement.To address the problems of low detection accuracy,high number of parameters,large computation and slow detection speed of most current deep learning models,YOLO v7 was used as the benchmark model,and Ghost Net,a lightweight detection network,was introduced into the YOLO v7 model as the network that extracts features,thereby reducing the number of parameters and computations of the model;meanwhile,to further decrease the computational complicatedness of the model and increase the model detection velocity,the depth-separable convolution rather than the common convolution was also substituted in the YOLO v7 feature fusion network.To strengthen the extraction of key features,decrease the weight of irrelevant information,and augment the feature representation capability of the network,the CBAM attention mechanism was integrated into the feature fusion network to extract key effective feature information;in addition,to address the possible overlap occlusion problem in detection,the K-means++algorithm was used to cluster anchor boxes in combination with microalgae microscopic image features,rather than using standard YOLO series 9 anchor box sizes,to achieve more rational and effective prediction boxes and further improve the model detection accuracy.Through ablation experiments and comparison tests with other models,the results showed that the proposed YOLO v7-MA model achieves the highest detection m AP of 98.56%on the microalgae dataset,and the quantity of parameters of the model was only 22.64×10~6,the FLOPs was38.45×10~9,and the FPS transmitted on the GPU was 51.86.The detection performance was better than the models Faster RCNN-VGG16,Faster RCNN-Resnet50,YOLO v4,YOLO v4-Mobilenet v3,YOLO v4-VGG16,YOLO v4-Resnet50,YOLO v5s,and YOLO v7.(3)Building an intelligent online detection platform for microalgae.Based on the YOLO v7-MA microalgae detection model proposed in this paper,an intelligent online detection platform for microalgae was built for use by non-computer related inspectors.The platform interface was based on the Streamlit application framework,which supported image detection,video detection,camera detection,and batch detection,etc.Users could select different detection models,set confidence threshold and Io U threshold,and directly displayed and downloaded the detection results.The platform could realize fast,accurate and efficient intelligent detection of microalgae,saving detection time and cost,significantly reducing the workload of inspectors,and having strong reference significance to the actual microalgae detection needs.