Fine-grained Image Classification Of Fruit Fly Via Convolutional Neural Network

In recent years,deep learning has developed rapidly,and a lot of achievements have been made in many research directions in the field of computer vision,such as image classification,image segmentation,etc.,which will bring significant breakthroughs in the field of agricultural informatization relying on subjective visual judgment.Taking fruit flies as an example,fruit flies are one of the important pests of horticultural crops in the Asian Pacific region.The identification of fruit flies is an important part of quarantine work.Establishing a fruit flies feature self-extraction classification and identification system is of great significance for quarantine work.The existing fruit fly classification task is the sample image of the thoracic-back region of the fruit fly or fruit fly in the HD shooting instrument,and the feature points of the region are selected by manually marking the feature,and finally the model is calculated by using the machine learning algorithm.The establishment of the feature marking process is cumbersome,and the professional requirements for the photographing instruments and users of the fruit fly samples are high,and the characteristics of the fruit flies cannot be automatically extracted.Based on the above problems,this paper proposes a convolutional neural network algorithm using deep learning algorithm to realize the classification model of fruit fly feature self-extraction.The main work includes:?1?Fruit fly image classification model based on convolutional neural network:In order to solve the problems such as few images of fruit fly samples,high cost of sample image acquisition equipment and tedious manual design and feature extraction process,a pre-training convolutional neural network vgg-16 model was proposed to complete the classification and recognition of fruit fly samples.Firstly,the clear fruit fly sample images were blurred by using gaussian fuzzy algorithm to reduce the clarity of the sample images.Secondly,the processed sample images were trained and classified by using the vgg-16model,and finally the classification model was obtained.The experiment shows that the model can automatically extract the whole characteristics of fruit fly for effective identification,and the overall identification accuracy of the Bactrocera dorsalis Hendel,Bactrocera cucuribitae,Bactrocera tau and Bactrocera scutellata is 88.33%,which has a good application prospect.At the same time,it solves the problem that the existing fruit fly classification and recognition system requires manual design and feature extraction in advance,reduces the requirements for the fruit fly sample shooting instrument,and improves the practicability of the model and the work efficiency of quarantine personnel.?2?Bilinear CNN-based fruit fly fine-grained image classification model:The similarities and differences between traditional image classification tasks and fruit fly image classification were analyzed in detail.It was found that the four types of fruit flies belonged to the fruit fly family,and the difference between the image types was small,so the fruit flies were Image classification tasks are combined with fine-grained image classification tasks.The discriminative information of fine-grained-level images often exists in very fine areas.It is not very reasonable to use the deep convolutional network method to extract the features of the entire image level directly in the above?1?.Therefore,it is proposed to use the mainstream fine-grained image weakly supervised classification.The model Bilinear CNN model classifies the fruit fly images.Experiments show that the model can effectively identify fruit flies,and the overall recognition accuracy of the Bactrocera dorsalis Hendel,Bactrocera cucuribitae,Bactrocera tau and Bactrocera scutellata is 91.66%,which is about 3%higher than the CNN model.The accuracy rate indicates that the fine-grained classification model has a certain improvement in performance compared with the CNN model in the fruit fly classification task,and has a good research prospect.?3?Feature-based fusion of FB-CNN fruit fly fine-grained image classification model:A detailed analysis of the characteristics of traditional fruit fly classification task extraction and its extraction method,it is found that artificial is usually extracted on local areas such as fruit fly wings and middle chest plate The features are classified and there is a relative positional relationship between the features on the fruit fly wing,and many studies have confirmed that the convolutional layer of the convolutional neural network can extract mid-level features^[22],and these mid-level features are often Corresponding to the local area of??the object,it is helpful to extract the characteristics of local areas such as the fruit fly wing and the middle chest back plate.Therefore,the method in?2?is improved,and the FB-CNN model with mid-level and high-level feature fusion is proposed to classify the fruit fly image.Experiments show that the model can effectively identify fruit flies,and the overall recognition accuracy of the Bactrocera dorsalis Hendel,Bactrocera cucuribitae,Bactrocera tau and Bactrocera scutellata is 96.67%,compared with CNN model and B-CNN model.The classification performance has been greatly improved,indicating that in the fruit fly classification task,combined with the mid-level feature can improve the classification performance,and it has certain reference significance for the research of other fine-grained image classification.The main contributions of this article are:?1?For the fruit fly image classification task,the fruit fly image classification task is divided into two tasks:traditional image classification and fine-grained image classification.The convolutional neural network model and the fine-grained image classification model B-CNN model are proposed for classification and improvement.The accuracy of the classification of fruit fly images.?2?Combining the traditional artificially extracted features,the FB-CNN model with mid-level and high-level feature fusion is proposed to classify the fruit fly images.The experimental results show that the method can improve the fruit fly image classification results.Compared with the Bilinear CNN and CNN models,the FB-CNN model has the best classification performance in the fruit fly image classification task.