Research And Applications On Agricultural Crop Pest Detection Techniques Based On Deep Learning

Nowadays,pest is one of the main causes leading to crop production failure in the agricultural field.The task that how to carry out professional and effective pest monitor-ing and prevention is significant for agricultural production.At present,most of the pest monitoring methods in China rely on "eye and hand inspection" from agricultural ex-perts,which are obviously time-consuming,labor-intensive,subjective and error-prone.In this case,researchers from all over the world have carried out numerous works on the the theoretical methods and systems of pest image detection and recognition that provide technical support for intelligent pest diagnosis and early warning.However,due to the unconstrained conditions in the field,the pest images meet the difficulties of inter-species similarity,scale variance,various postures,lighting effects,crop oc-clusion and so on,which result in low identification accuracy and weak generalization in applications.Comparing with these approaches,deep learning techniques that have emerged in recent years are widely used in the field of computer vision area including pest detection task.In addition,they also have performed well in a series of large-scale and fine-grained recognition tasks,and thus have gradually become a preferred solution for agricultural pest detection.This dissertation will systematically investigate and study the theories and methods for pest detection task based on deep learning,including the novel convolutional neural network architecture and optimization methods for pest detection framework as well as explore the new methods for applicable and effective for pest monitoring in the field.Furthermore,we will also analyze the limitations of current generic object detection algorithms and propose a novel alternative.The major contributions of this dissertation are as follow:1.Construct two standard datasets of large-scale pest images.Through the accumu-lation of the past 5 years,two standard pest detection datasets are constructed from the trap and in-field environments corresponding to the different application scenarios.In order to meet the requirements of pest detection task,professional pest object annotations are performed on the collected large-scale pest images.Besides,considering the difficulties of practical pest detection task,the whole datasets are in-depth analyzed and constructed by our designed pest collection equipment.Finally,we constructed Multi-class Pest Dataset 2018(MPD2018)and AgriPest in two different environments of crop pest target detection datasets,containing 88K and 49K pest images,with 582K and 264K pest annotations re-spectively,that are verified the validity and applicability by benchmark experi-ments.2.Propose a hybrid global and local feature based pest detection method.This ap-proach is built on the two-stage object detection architecture to improve pest fea-ture representation.Specifically,we extract high-quality global features for pest images by designing the global feature pyramid network containing channel ac-tivation module and spatial activation module.In addition,the enhanced local features of pest objects are extracted by local activated region proposal network,which is implemented by the collaboration of contextual information as well as self-attention mechanism.Finally,the deep features of pest are fused from global and local features to realize accurate pest detection and recognition.This method achieves the state-of-the-art performance on our MPD2018 and AgriPest datasets and further verifies the effectiveness of the proposed algorithm on crop pest de-tection task.3.Propose a generic object detection method based on scale-aware feature selective module.In this approach,we firstly present a RoI feature pyramid structure to extract local features on each feature levels.Subsequently,in order to achieve scale-aware adaptation between hierarchical features,a trainable weighting gate function is designed to automatically learn the weights of various feature levels for the input instance to realize scale-aware feature selection and fusion.Experi-ments show that the proposed method achieve the state-of-the-art object detection performance on the large-scale dataset MS COCO and also performs well in our pest detection task.