Convolutional neural network segmentation for counting polyps in images

V nalogi naslavljamo problem detekcije polipov meduz na slikah ostrig. Moderne metode detekcije objektov so pogosto sestavljene iz dveh faz. Najprej se na potencialnih lokacijah generirajo hipotetične regije, nato pa se posamezno hipotetično regijo klasificira v pripadajoči razred, glede na objekt, ki ga regija vsebuje. V delu se osredotočamo na drugačen pristop k detekciji objektov, saj najprej z uporabo konvolucijske nevronske mreže generiramo segmentacijsko masko objektov na sliki, nato pa z interpretacijo maske dobimo natančno lokacijo iskanih objektov. Razvito metodo SegCo uporabljamo za reševanje problema detekcije polipov meduz na slikah školjk. Rezultate predlagane metode smo primerjali z rezultati modernih metod detekcije objektov. Rezultate smo primerjali z modernim učljivim detektorjem RetinaNet in specializirano metodo za detekcijo polipov PoCo. V primerjavi z detektorjem RetinaNet, metoda SegCo dosega 2% izboljšavo mere F-1, v primerjavi z detektorjem PoCo pa 24% izboljšavo. V sklopu naloge je bil razvit tudi program, ki z uporabo predlagane metode omogoča avtomatsko detekcijo in štetje objektov na slikah. Delovanje programa ni omejeno le na detekcijo polipov, temveč z učinkovitim uporabniškim vmesnikom omogoča učenje novih modelov predlagane metode, zaradi česar lahko program detektira objekte tudi na drugih vrstah slik.We address the problem of jellyfish polyp detection on images of oysters. Modern methods of object detection often utilize convolutional neural networks for feature extraction and work in two stages. First, hypothetical regions are proposed at potential locations, the features of the regions are extracted and are later classified according to the object they contain. In this work we focus on an alternative aproach to object detection in which we first use a convolutional neural network to obtain an image segmentation mask which we then interpret to extract the precise location and shape of the objects in the image. We use the proposed method SegCo to address the problem of jellyfish polyp detection on images of oysters. We compare the results of the proposed method with current state of the art object detecion methods. We compare the results the state of the art learnable detector RetinaNet and the specialized polyp detection method PoCo. In comparison with RetinaNet, SegCo achieves a 2% improvement in F-1 score and in comparison with PoCo, the achieved improvement is 24%. In addition we developed a program, which utilizes the proposed method to enable the user to automatically detect and count objects in images. The application of our program is not limited to the detection of jellyfish polyps, as it contains an efficient user interface for training new models using the proposed method, which enables the user to easily apply our method to objects in other types of images