Plant fine granularity recognition method based on bilinear convolutional neural network

Abstract

The invention discloses a plant fine granularity recognition method based on a bilinear convolutional neural network. The method comprises the following steps: S1, carrying out the coarse granularity feature extraction through a Relu activation function, and obtaining a coarse granularity feature X1; s2, obtaining the overall representation of the bilinear model; s3, in the step S2, performing bilinear fusion on features extracted by two feature extraction functions A and B in the bilinear model to obtain a matrix b; s4, performing summation pooling on the matrix b to obtain a matrix xi, and performing multi-dimensional vector expansion on the xi to obtain a feature vector x; s5, moment normalization and L2 normalization are carried out on the feature vector x to obtain bilinear features, the structure is scientific and reasonable, use is safe and convenient, the bilinear model is used for effectively extracting subtle differences between high-throughput plant expression species and even between individuals, the feature extraction advantage of bilinear pooling on fine-grained images is fully played, and the method is suitable for large-scale popularization and application. And relatively high plant fine granularity recognition precision is achieved.

Description

technical field [0001] The invention relates to the technical field of plant identification algorithm improvement, in particular to a plant fine-grained identification method based on a bilinear convolutional neural network. Background technique [0002] The traditional plant identification method mainly relies on manual observation and measurement to analyze the appearance, texture, color and other morphological phenotypes of plants. This method has problems such as low work efficiency and low accuracy during measurement. The development of learning in the field of computer vision has made deep convolutional neural network (CNN) processing of two-dimensional natural scene images one of the research focuses of domestic and international attention. The majority of agriculture and forestry researchers will also use CNN as the representative. The application of deep learning technology to plant phenotype research has opened the era of intelligent research on plant phenotypes. R...

AI Technical Summary

Problems solved by technology

[0003] 1. The above method does not specifically analyze the plant image, thus ignoring the unique characteristics of the plant image. There are external interference factors such as posture, illumination, occlusion, and complex background in the data collection, resulting in similar images. The different samples of different samples are very different, and there are growth variations in plants, which further lead to obvious differences in the representation of different stages of plant growth, forming the fine-grained characteristics of "intra-class differences";

[0004] 2. Plants have a detailed division of biological subcategories or subcategories, and there are certain biological similarities between each subcategory, which leads to the fine-grained recognition problem of "similarity between categories". The existing CNN-based recognition model is applied to this Plant recognition with fine-grained characteristics cannot achieve optimal recognition results, resulting in technical problems such as low accuracy and generalization robustness in actual recognition;

[0005] 3. Most of the relevant research at home and abroad is only applicable to the coarse-grained identification of plants with large intermediate differences, which cannot effectively solve the problem of fine-grained identification caused by direct "interspecies similarity", and cannot meet the needs of refined agriculture;

Images