Feature Detection Method of Small Sample Poultry Egg Image Based on Prototypical Network

Machine vision has developed into a mainstream testing method in the field of nondestructive testing of poultry eggs due to its advantages such as high detection speed, high stability and low cost. A large number of egg images are often used as data support to achieve better detection results. However, the collection cost of egg image data is relatively high, and it costs a lot of manpower and material resources. Therefore, it is hoped to find a method similar to face recognition for small sample egg image detection. To solve this problem, a prototypical network suitable for the detection of small sample egg images was proposed. The network used the inverse residual structure of attention-introducing mechanism to build a convolutional neural network to map different types of egg images to the embedded space, and Euclidean distance measurement was used to test the types of egg images in the embedded space, so as to complete the classification of egg images. The network was used to verify the classification detection effect of fertilized egg and unfertilized egg, double yolk egg and single yolk egg, cracked egg and normal egg under the condition of small sample. Its detection accuracy was 95%, 98%, 88%, respectively. The test results showed that the method effectively solved the problem of insufficient samples in the detection of poultry egg image, and provided an idea for the research of nondestructive detection of poultry egg image. In future nondestructive testing of poultry egg images, a small amount of poultry egg images can be collected to achieve better detection results. © 2021, Chinese Society of Agricultural Machinery. All right reserved.