Estimating weight from diyarbakir region watermelon images using deep learning methods

In this thesis study, using the images of watermelons grown in the Diyarbakır region, the weight estimation of watermelons was made by deep learning methods. 5000 watermelon images were used in the study. Photographs of watermelons grown in the Diyarbakir region were taken in markets and neighborhood markets and recorded in a computer environment. After the backgrounds of the collected watermelon images were taken, their masks were made in the Python program. Masks are included in the training file to be used in the U-Net architecture. Convolutional neural networks and U-Net architecture are used in deep learning methods. Segmentation of images has been done successfully. The U-Net architecture has successfully predicted the watermelon image geometrically at a rate of 99.65%. The area ratio of the watermelon in the image was calculated by using the pixel area method from the images obtained in the U-Net model. After the height and width pixels of the images were determined, training with the artificial neural network was started. To find the best architecture, artificial neural networks were trained with 9 different architectures. The best architecture was the one that was divided into 1024 units with 4 hidden layers and ReLU was used as the activation function. By training the watermelon data obtained from the images in multilayer artificial neural networks, the test accuracy rate was 92.59% and the training accuracy rate was 94.43%. As a result, thanks to the deep learning method created, a program that will estimate the weight of watermelon images taken at a distance of 65 cm will contribute to the studies in the field of digital agriculture.