Research On Object Detection And Image Segmentation Methods Of Plateau Pika Based On CNN

Plateau pika is endemic to the Qinghai-Tibet Plateau.Plateau pikas,while participating in maintaining the ecological balance of the plateau,are prone to flood and form rodents,causing economic losses to herders.Therefore,it is of great significance to correctly predict the population of plateau pika and take corresponding control measures to reduce the occurrence of rodent harm and maintain the ecological balance of plateau grassland.With the rapid development of technology in the field of machine vision,it has become an achievable goal to obtain the population of plateau pika in the region through intelligent monitoring equipment.The detection and segmentation of targets in plateau pika images is the basis and core for obtaining statistical data.The plateau pika images have the characteristics of small target size,complex background,and insignificant target features,which makes the plateau pika images difficult to detect and segment.Based on the Faster-RCNN model,this paper improves the Faster-RCNN model based on the characteristics of plateau pika images to achieve more accurate detection of plateau pika.The Faster-RCNN model is a two-stage detection model.In order to solve the problem of repeated sampling caused by the anchor box introduced in the first stage and the problem of low accuracy of small target detection caused by background noise brought by the sampling of the anchor box,the paper proposed Method for generating regions of interest based on semantic segmentation.This method converts the search problem of the region of interest into a binary semantic segmentation problem of the foreground and background.Then the segmentation mask obtained by the semantic segmentation is used to directly obtain the position and size of the region of interest through median filtering and connected domain analysis.Therefore,the introduction of anchor points is avoided,thereby simplifying the model while reducing the computational complexity of the model and reducing the impact of background noise.Experiments show that based on the plateau pika images in natural scenes,the region of interest generation method based on semantic segmentation in this paper is 27.75 higher than the RPN network used in the first stage of the Faster-RCNN model.%,Better detection performance.The method of generating a region of interest based on semantic segmentation can train a semantic segmentation model by detecting labeled data,and then obtain a preliminary segmentation result of the target.Based on this,a small sample semantic segmentation model based on weakly supervised transfer learning is proposed for the characteristics of small amount of data and limited segmentation label data when training the semantic segmentation model in engineering practice.In this paper,the target detection data set is used as a weakly supervised training data set.First,use this data set to train a semantic segmentation model to achieve a mapping from the original image to the coarse segmentation mask.In the second stage,this paper proposes a foreground-based the background expansion method of data separation can expand the extremely limited training data into a relatively larger training set without requiring additional training.On this basis,the extended segmentation training data is used to fine-tune the model trained in the first stage to obtain the final semantic segmentation model.Using this method,the target detection data can be used to reduce the dependence on the segmentation label data in the training of the semantic segmentation model,and at the same time improve the model’s efficiency in using limited segmented samples and the model’s segmentation performance.The experiments show that the segmentation performance of the model in this paper is close to that of the semantic segmentation model trained with a large number of segmented datasets,and it is better than that of the Chan-Vese model.In addition,the experiments show that the data augmentation method based on foreground and background separation proposed in this paper can effectively improve the segmentation performance of the model.