Research On Microscopic Images Of Plant Fragment Detection And Recognition Method Based On Deep Learning

Fecal microscopic analysis is a widely used method for feeding analysis of ungulates in the field,revealing the interrelationships between animals and their habitats.However,this method requires manual identification by researchers,which is susceptible to subjective factors such as operator experience and is time-consuming and labor-intensive.Therefore,it is important to investigate the use of artificial intelligence techniques to achieve efficient detection and identification of plant fragments in the feces of ungulates to assist in the predation analysis of wildlife.The main research contents include:(1)To address the problem of poor detection accuracy due to the size difference of plant fragments,a microscopic image target detection algorithm based on multi-scale feature fusion of plant fragments is studied.The YOLOv5 target detection framework is adopted,and the weighted bidirectional feature pyramid network BiFPN is applied to improve the neck network to enhance the algorithm’s ability to fuse multi-scale features of plant fragments;the Alpha-IoU loss function with added weight coefficients is introduced to improve the localization effect of borders.The experimental results show that the improved detection algorithm achieves the purpose of improving the detection effect of plant fragments microscopic images,and the average detection accuracy is improved by 2.9%.(2)To address the problem of low recognition rate due to blurred edge features of plant fragments,we study the microscopic image recognition algorithm of plant fragments based on local and global feature extraction strategies.The local feature extraction channel is constructed using EfficientNetV2 network to make full use of the local fine-grained features in plant fragments;the MBConv block is improved by using CBAM module with added spatial attention,and the global feature extraction channel is constructed by combining multi-axis attention mechanism to form a C-Max Transformer module to enhance the acquisition of global Contextual information is acquired;the integration mechanism is used to fuse the two channels to improve the classification performance.The experimental results show that the algorithm using local and global feature extraction-based strategies has better recognition effect and the recognition accuracy reaches 88.3%.(3)To address the problem that labeled plant fragment microscopic images are too expensive to label and have a large number of redundant unlabeled images that are not fully utilized,a self-supervised contrast learning-based microscopic image recognition method for plant fragments is investigated.An asymmetric momentum contrast learning architecture is used to pre-train the encoder using unlabeled data;a combined data augmentation means is used to construct an agent task to enhance the feature extraction capability of the encoder;the fine-tuning process is optimized using a feature distillation method to make the features learned by the pre-trained model more easily migrate to the downstream classification task.The experimental results show that the recognition method based on self-supervised contrast learning can effectively utilize unlabeled data to improve recognition,and the recognition accuracy of plant fragments reaches 91.2%.