In order to meet the non-contact non-destructive testing requirements of composite materials, the composite defect (taking the pores as an example) detection technology was studied, and a method based on machine vision for composite defect image recognition was proposed. The method adopts the method of adaptive histogram equalization to improve the image contrast, enhance the image detail, and modify the deep convolutional neural network Yolo-V3 based on the digital ray technology for composite defect image acquisition and the contrast of the contrast of the ray image. The input and output are trained by self-built data sets to achieve high-precision identification of composite defects, and the calculated average accuracy is 86. 04%.