Research On Image Recognition Of Metamorphosed Insects

Field insect recognition plays a key role in the protection of agroforestry crops.Large scale insect identification in natural scenes is extremely challenging due to the characteristics of large intra-class and small inter-class differences.Therefore,accurate and efficient detection of insects in the natural environment is very important for pest control and improving the economic benefits of agroforestry.Most traditional insect identification methods rely on the professional knowledge of entomologists.These methods not only have the disadvantages of strong subjectivity and low efficiency,but also ignore the impact of insects’own metamorphosis and development on the identification task,which cannot meet the requirements of current agricultural intelligence and information.In order to achieve a higher level of insect image recognition,this study used convolutional neural network to extract image features and carried out the following studies:(1)A new data set was built based on the IP 102 data set.The problems of small sample size and single image background exist in most of the existing insect image data sets.In order to meet the needs of deep learning network,semisupervised method and the YOLO algorithm were used to process the image samples in IP 102.A new insect image dataset(IP102-YOLO)was constructed with the processed images.97,494 images,covering 102 types of common insects were contained in IP 102-YOLO,which not only met the research content of this paper but also provided support for insect image recognition research based on deep learning.IP 102-YOLO is more suitable for insect image recognition tasks than IP 102.When the same algorithm is used,the average accuracy of the former is improved by 11.575%.(2)The two-stage insect image recognition model(TIR)is proposed in this paper,which can alleviate the interference of insect metamorphosis development characteristics on the image features of deep neural network learning by means of clustering before recognition.The performance of the TIR model is excellent,and the recognition accuracy is 83.64%,as demonstrated by the experiment.Furthermore,any convolutional upgrade network can replace the backbone network of TIR based on task requirements.Experiments have shown that the accuracy of image recognition can be effectively improved through the combination of TIR and traditional network.(3)The recognition capability of the original TIR is improved by focusing on the important information of the image through an attention mechanism in an improved algorithm for insect image recognition.By adding channel attention,spatial attention and fractional attention to the TIR model,each attention mechanism was proven by the experiment to be capable of improving the model’s image recognition ability.When three kinds of attention were introduced into the model at the same time,the accuracy of image recognition was increased by 2.64%to 86.38%.(4)An insect image recognition system is designed and developed in this paper based on the above research.The proposed system mainly has two functional modules,namely insect image recognition and insect knowledge encyclopedia.Single object images can be recognized and entomology knowledge can be learned by users using this system.It was shown by the test that the system can meet the demand of image recognition well and good performance is exhibited in terms of running speed and recognition accuracy.