Research And Test On Plant Leaves Recognition System Based On Deep Learning And Support Vector Machine

Plant identification is the basis of plant protection and utilization. With the development of image recognition technology, the use of image processing technology of plant leave characteristics, such as shape and texture, can greatly improve the identification efficiency, realize automatic and quickly classification and identification of plant. This paper designs and develops the identification system of plant leaves via VGG-16 (Imagenet-vgg-verydeep-16)、Matlab SVM (Support vector machine) toolkit and Matlab GUI (Graphical user interface), constructs plant leaf photo databases of natural and special environments and analyzed the influence of leaf angle, background, photographic sensitivity, color temperature and blade damage on the results of plant identification. Through this study, this method is proved be good in plant identification, which provide the base for the development of plant leaf image recognition technology in the future. The research results are as follows:① Based on software requirements, we set system environment, designed system flow and system interface, wrote system code and realized the function of the training & identification of plant leaves with total 1020 photos under natural environments of 34 plant species and 735 photos of 5 plants at specific environments.② More training photos, there were higher identification rate of plant leaves. In 20 photos of every plant species, when 5,10 and 15 photos were selected randomly for training, the average identification rate of 10 test photos of each species were 64.65%、 78.12% and 89.06% respectively, and the difference of identification rate among test sets was significant. When taking all 20 photos of leaves of each plant species as training set, the test rate of remaining 10 photos was 93.24%.③ Influence of the environmental factors on identification rate is different. The influence of large angle of leaves and low photographic sensitivity is significant. In the test to five plants of 0,30°and 60°angle, the identification rates were 100%,100% and 60% respectively; In the photo shooting test of 8 background colors and three different background textures, the all identification rates were 100%. At the photographic sensitivities of camera were ISO400, ISO800, ISO1600 and ISO3200, the identification rates were 89%,100%,100% and 100% respectively; The plant identification rates of photos in color temperature from 3000 to 9000K were all 100%; when the damage area was not exceeding 10% of the total leave area, the identification rate is 100%.