A hybrid model based on the convolutional neural network model and artificial bee colony or particle swarm optimization-based iterative thresholding for the detection of bruised apples

In this study, apple images taken with near-infrared (NIR) cameras were classified as bruised and healthy objects using iterative thresholding approaches based on artificial bee colony (ABC) and particle swarm optimization (PSO) algorithms supported by a convolutional neural network (CNN) deep learning model. The proposed model includes the following stages: image acquisition, image preprocessing, the segmentation of anatomical regions (stem-calyx regions) to be discarded, the detection of bruised areas on the apple images, and their classification. For this aim, by using the image acquisition platform with a NIR camera, a total of 1200 images at 6 different angles were taken from 200 apples, of which 100 were bruised and 100 healthy. 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 preprocessing 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. Second, the threshold value was determined by means of the 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. Finally, the apple images were classified as bruised and healthy objects by using this threshold value. In order to illustrate the classification success of our approaches, the same classification experiments were reimplemented by directly using the CNN model alone on the preprocessed images with no ABC and PSO approaches. Experimental results showed that the hybrid model proposed in this paper was more successful than the CNN model in which ABC-and PSO-based iterative threshold approaches were not used.