Research On Intelligent Detection Algorithm Of Crop Pest Image

Integrated pest management is an important measure to maintain high crop production and is an essential part of intelligent plant protection.Currently,vision-based automatic crop pest recognition and detection technologies have been widely used to improve the efficiency of plant protection personnel and reduce objective judgments brought by humans.With the development of artificial intelligence,common object detection tasks have achieved great success.Thus,researchers applied it to address crop pest detection tasks,and it has also become a hot topic in the field of computer vision in recent years.However,there are still challenges in crop pest image recognition and detection in practical applications,and the difficulties of small-scale,multi-scale,and high similarity in appearance of crop pest images lead to unsatisfactory detection accuracy.To address the main difficulties of pest image detection,this dissertation focuses on four key challenges to develop detection algorithms:small scale of crop pests,multi-scale crop pests,small differences in appearance between crop pest classes,and improved detection speed.The main work of this dissertation can be summarized as follows:1.A small-scale pest detection method based on channel recalibration feature pyramid network is proposed.Since small-scale targets occupy few pixels,resulting in insufficient representation information,it is more difficult to learn effective features during the training process.To address the problem of low accuracy of small-scale pest detection,two key components,the channel recalibration feature pyramid network and the adaptive anchor module,are proposed.The channel recalibration feature pyramid network can capture discriminative features and learn finer object features,significantly improving the recognition accuracy of small target pests,while the adaptive anchor module is used to correct the problem that small-scale pests are prone to inefficient matching of anchor boxes and real labeled boxes.This strategy effectively reduces the search range among tiny pests and improves the accuracy of regression bounding boxes in the refinement network.The experiment results on the LMPD2020 dataset well demonstrate the superiority of our proposed method.2.A scale-aware network-based pest detection method is proposed.For the problem of accurate detection of multi-scale crop pests,the proposed method consists of three key components,namely,a high-level semantic feature extraction module,a lowlevel feature enhancement module,and a dynamic scale-aware head network.The highlevel semantic feature extraction module can retain the rich information of high-level features and help to build the feature pyramid network.The low-level feature enhancement module optimizes the low-level integrated feature map and provides more fine-grained feature information.In addition,the dynamic scale-aware detection head network can improve the detection performance of small targets by adaptively selecting the appropriate detection field according to different object scales.The comparative experimental results on APHID-4K and LMPD2020 datasets show that our proposed method achieves the best detection results in the same period.3.A pest detection network based on discriminative feature fusion is proposed.In the task of crop pest image detection,it is necessary to distinguish those objects belonging to which pest specific categories.At the same time,pests belonging to the same family are highly similar in appearance,which usually require expert plant protection personnel to recognize them accurately.To address the problem of small differences in appearance between pest categories,the pest detection network with discriminative feature fusion proposed in this paper first utilizes multi-scale discriminative feature fusion pyramids to fuse pest information from multi-scale features through an adaptive channel fusion module and a global context module.Meanwhile,an adaptive feature region proposal network is designed to extract pest features of interest more accurately by solving the problem of anchor box and feature misalignment during the iteration of the region proposal network.Extensive experiments on the LMPD2020 dataset show that our proposed method achieves excellent detection performance.4.An attention-based single-stage pest detection network is proposed.For crop pest recognition and detection in practical applications requiring high efficiency of detection models,the model time complexity is reduced under the condition of guaranteed accuracy.The proposed model uses a dynamic training sample selection algorithm in the training phase to select high-quality positive samples to improve the quality of model training.It dynamically captures high-quality training samples that contain multi-scale background information,thus reducing the problem of false-positive samples easily caused by small-scale targets.Then,an attention-based dynamic detection head network for obtaining highly representative semantic features,which improves the ability of location and recognition pest objects.Through extensive comparison experiments on the CropPest24 and MPD2018 datasets,it is shown that our proposed method ensures accuracy while still having low model FLOPs and achieves state-of-the-art performance.