Image Classification Of Crosswise Vegetable Pests Based On Multi-scale Attention Network

Cruciferous vegetables are always the largest vegetable planted in our country,so the control of pests is particularly important.Pest detection technology is a key part of pest control.The current deep learning method is weak in the research of cruciferous vegetable pest image,which is difficult to meet the requirements of vegetable seed industry.In order to improve the identification efficiency of cruciferous vegetable pests,this paper constructed the data set of insect pest images.The Resnet50 model was used as the basic network to improve the traditional attention network and multi-scale convolution algorithm optimization model,and the label smoothing method was used to improve the traditional cross entropy loss function to build the cruciferous vegetable pest detection model.The specific work of this paper is as follows:(1)Aiming at the problem of image channel and spatial information loss caused by the increase of model depth,a method of weighted image key features by hybrid attention network was proposed.In the process of image feature aggregation,the proportion of important features in channel and space was increased to retain more classification information.However,the introduction of traditional attention network will inevitably increase the number of excessive parameters.Therefore,by constructing lightweight attention network in this paper,we can not only learn more key features of images,but also do not need to increase the calculation amount of the model too much.(2)An improved multi-scale convolution algorithm is proposed to gather different visual information of images,aiming at the problem that the feature loss of the front layer network of the model leads to the reduction of the feature amount of the back layer network.By setting the expansion rate of a set of cycles,the multi-scale convolution algorithm in this paper can effectively increase the number of receptive fields in the image,obtain pest feature information of different scales,and learn more important pest features for training on the three-channel feature map input at the first layer of the model.(3)Aiming at the problem of poor fitting effect of model data,a method is proposed to improve the traditional cross entropy loss function by using label smoothing idea.By introducing label smoothing into the traditional cross entropy loss function,changing the value range and expression of the conventional label,and reducing the error between the predicted value and the new label value,the generalization ability of the model can be effectively improved.In this paper,the improved multi-scale attention network was used to classify ten major pests of cruciferous vegetables.The classification accuracy reached 94.45%,7.81%higher than that of Resnet50 model,which can effectively improve the identification effect of pests.