A Flower-Image Search System Based on Deep Learning

摘要

#$%^&*AU2019100352A420190509.pdf#####ABSTRACT The invention patent specifically designs a flower image recognition system based on deep learning and neural network learning. First, images can be acquired by web crawler from the Internet as many as possible. Second, the searched images are divided into 2 parts: the train section and the test section. The train section is used for training deep learning architecture while the test section is used for test performance. Third, the train section will be input into the system which contains several layers of convolutional neural networks. By adjusting the parameters from the neural learning network such as learning rate, drop-out rate, and decay rate, the system will finally reach the optimal performance. In this case, users simply upload an image and the system will input it into corresponding structures based on the features of the image. Finally, the system will automatically return some similar images to the users. This invention can be applied to other image recognition areas such as mood-detecting and vehicle recognition. 1image Cloud base32 to storage - to base32 platform image Server-side Client-side Figure 12 STRT)DROPOUT OPTIMZR ORIGINAL DATA P RFDTCIOIA ReZATIK1N (BCR ' WFR) PRDIC OCUAC SAMPLE TOPTIMLUTNI N ENLA.RGEMENT SE(75) TEST SRT(6839) RENZ T AXPOOLING NVLTINONOUTO ENCODING NORMAILIZATIO)N OF THE CRAY VALUE 7 t I FU LUFULL RESHAPE THE -- + ~~~TRANSFORM 3 CHANNELS TO I CHNE CONCED - CNNCE 4CO IMAGE MAYR1X LAE1LAYER2 nFGIuIAnIizXIION OPTIMAL PARAMNETER CL Figure 13 8