Research On Pest Detection Method Based On Unbalanced Data

Agriculture is the foundation of national economic development and the root of social stability and unity.However,the continuous high and heavy occurrence of agricultural pests directly affects national food production and security.Accurate and timely identification and detection of agricultural pests is the key to enhancing pest monitoring and early warning.In recent years,with the remarkable progress of convolutional neural networks,researchers utilize deep learning-based object detection techniques for the pest detection task to solve deficiencies of traditional machine learning-based methods,such as insufficient generalization ability and performance.However,affected by the properties of pests,the acquired pest data has serious imbalanced problems that bring great challenges to the accuracy and speed of detection algorithms.We study key techniques and algorithms to improve the performance and efficiency of pest detection based on existing research,which mainly involves three perspectives:pest category imbalance,scale imbalance,and distribution imbalance.The main work and innovations of the dissertation are as follows:1.To address the pest category imbalance problem,this study proposes a pest virtual image expansion method based on the generative adversarial network.The method uses the designed Auto-grabcut algorithm to automatically extract single pest images from the deep learning training dataset and filter their background pixels,obtain highresolution virtual pest images through a generative adversarial network to enhance the total information entropy of the single pest dataset,and images from the single pest dataset are fused with blank background images to generate pest virtual images according to certain rules for the pre-training process of the detection network,thus improving the detection accuracy of few-shot pest categories.The LLPD-26 large-scale multiclass pest dataset was constructed to verify the effectiveness of the proposed method.Extensive experiments show that the method proposed in this study can improve the performance of the detector by enhancing the attention of the convolutional neural network for few-shot pest categories.2.To address the scale imbalance of pests,this study proposes a pest detection network based on a multi-scale super-resolution feature enhancement module,which improves the deficiency that the data expansion method cannot solve the scale imbalance problem.The designed pest detection algorithm mainly consists of two core designs:the multi-scale super-resolution feature enhancement module and the Soft-IoU calculation method.The multi-scale super-resolution feature enhancement module can effectively solve the information imbalance at the feature level and improve the detection accuracy of small-size and multi-scale pests.The Soft-IoU calculation method improves the applicability of the detector for pest detection tasks by flexibly adjusting the IoU value.Extensive experiments have verified the validity of this study,and ablation experiments have shown that the multi-scale super-resolution module can be combined with different detectors in a plug-and-play manner and enhance their detection performance.3.For the distribution imbalance of pests,this study proposes a pest detection method based on a Transformer feature pyramid network,which improves the shortcomings of light-trapping pest detection methods in dense distribution problems.The method includes two core designs:a Transformer-based feature pyramid network and a multi-resolution training method.The Transformer-based feature pyramid network improves the feature extraction capability of tiny-sized dense distribution pests.The multi-resolution training method is used to solve the multi-view problem caused by distribution imbalance,thereby improving the detection performance of fuzzy pests.In addition,due to the coarse-to-fine training pattern,the multi-resolution training method can simultaneously improve the performance and efficiency of detection networks.Due to the problem of distribution imbalance being the most prominent,the APHID-4K dataset was constructed by collecting aphid images to verify the effectiveness of the method.Abundant experiments show that our pest detection network obtains state-ofthe-art performance and has sufficient practical application value.