The detection of bruises on apples using machine learning and image processing

In this thesis, the apple images taken by using the image acquisition platform with near-infrared (NIR) camera were classified as bruised and healthy using a hybrid model based on artificial bee colony (ABC) and particle swarm optimization (PSO) iterative thresholding algorithms supported by a convolutional neural network (CNN) model. The proposed model includes image acquisition, image pre-processing, the segmentation of anatomical regions to be discarded and the detection of bruised areas on the apple images. A total of 1200 images at different angles were taken from 200 apples those of which were 100 bruised and 100 healthy by using the image acquisition platform. In order to increase the success of detection and classification, adaptive histogram equalization (AHE), edge detection and morphological operations were applied to the images in the pre-processing stage, respectively. First, in order to segment and discard the stem-calyx anatomical regions of the images, the CNN model was trained by using the preprocessed images. With this model, the stem-calyx regions were removed from the image. Second, the threshold value was determined by means of ABC/PSO-based iterative thresholding approach on the images whose stem-calyx regions were discarded, and then the bruised areas on the images with no stem-calyx anatomical regions were detected by using the determined threshold value. In order to illustrate the classification success of the proposed approaches, the same classification experiments were reimplemented by directly using Faster R-CNN model alone on the preprocessed images with no ABC and PSO approaches. Experimental results showed that this hybrid model proposed in the thesis was more successful than the Faster R-CNN model in which ABC and PSO-based iterative threshold approaches were not used.