Development of AI based applications to support insect breeding experiments for the circular economy.

To solve the time-consuming and laborious manual counting investigation and low recognition rate of larvae in the laboratory. This thesis studies the latest Flask framework and artificial intelligence-based image processing technology to realize a mobile application automatically classifying and counting larvae. In order to realize the counting and classification of larvae images for small data set, larva image segmentation based on grayscale thresholding and edge detection methods and an image recognition model based on transfer learning is proposed. Based on the VGG-16 model, a new fully-connected layer module was designed. The VGG-16 model was migrated to the model in the trained convolution layer of the ImageNet image data set. The collected image data set was divided into a training set, testing set, and validation set. In order to expand the data set of the image, the original set of the training set was rotated, flipped, and the like. Based on the training set after the expansion, compare two transfer learning methods, only trains the full connection layer, trains all the layers(convolution layer + full connected layer). The results showed that transfer learning trains model all layers can improve the recognition ability of the model. Compared with the new learning, transfer learning can significantly improve the model's convergence speed and recognition ability. Embed the trained model into the Flask framework to complete application functions. This application helps the experimenter reduce the workload, promotes biological science development, and provides a reference for image processing to identify tiny insects.