| Object detection,as a fundamental topic in the field of computer vision,mainly involves the classification and localization of objects.It has a wide range of applications in practical scenarios such as industrial defect,medical diagnosis,and agricultural pest detection.However,object detection still presents many challenges and opportunities for development while facing increasingly complex real-world situations.Therefore,in-depth research is helpful for promoting the advancement of both academic and industrial applications in this area of deep learning.With the widespread application of deep convolutional neural networks in the field of computer vision,it has gradually replaced the traditional manual feature design approach.Convolutional neural networks are able to learn abstract high-level semantic features and have good generalization effects in large-scale manually labeled image tasks.In object detection tasks,although convolutional neural networks can correctly extract features of regions of interest,there are problems such as scale inconsistency,complex backgrounds,object adhesion,and object occlusion in real image scenes,which weakens the generalization ability and reduces the detection accuracy of object detection algorithms.In addition,the lack of standardized datasets in specific fields(such as agricultural pests)also restricts the development of computer vision in these areas.This dissertation focuses on the issues related to receptive fields,feature extraction,sample matching,and specific field datasets,and explores a small-scale object detection model based on receptive field theory to propose a more effective solution for agricultural pest detection.The work of this dissertation includes the following aspects:1.In response to the difficulties in convergence,insufficient training and imbalance in the training process of object detectors,this dissertation proposes an enhanced representation learning method that draws on the working principles of the human visual system.By simulating the effective receptive field mechanism,it generates the positive samples required by the object detector.In addition,it is found that the size of the effective receptive field region is heavily influenced by the depth of the convolutional layer.Therefore,this dissertation fits the formula for the effective receptive field radius of different depth convolutional layers.This method enhances the object detector’s ability to learn subtle features.2.This dissertation proposes a positive sample filtering method based on effective receptive field that relies on region generation networks to filter candidate boxes to narrow the gap between the theoretical receptive field and the effective receptive field,thereby improving the representation learning of the object detector and balancing the positive and negative sample ratio in the two-stage object detector.Moreover,this method can be further extended to feature pyramid networks to address the quality of positive samples in different scale feature maps.Experimental results show that the proposed method can significantly improve the detection accuracy of the object detector in two large-scale generic datasets.3.This dissertation proposes a multi-input single-output object detection.framework aimed at solving the problem of high computational cost due to repeated feature extraction and deep network fusion in object detection techniques.Unlike traditional multi-input multi-output(Feature Pyramid Network,FPN)object detectors,this framework has the advantages of reducing model complexity and inference time overhead.Meanwhile,under this framework,this dissertation proposes a new dynamic positive sample selection strategy,which is designed based on the representation learning of the effective receptive field.This strategy realizes more concise and efficient feature extraction.4.In response to the lack of standardized object detection datasets in the field of agricultural pest control,this dissertation establishes an agricultural pest dataset named Pest24.The dataset has the characteristics of large-scale,small size,dense distribution,and high adhesiveness.This dissertation analyzes in detail the impact of multiple factors in the Pest24 dataset on object detection difficulty and tests the dataset with generic object detectors and the methods proposed in this dissertation to verify its universality.In addition,experimental results show that the proposed method achieves the optimal detection accuracy on this dataset,providing a solution for pest detection in the field of agricultural pests. |
