In the segmentation process, it is necessary to check the similarity of the points that make up the image, and the similar points that are adjacent to each other are placed in a single area.
In the two-stage form, it is necessary to first form a feature vector at each point and then put the feature vectors into clusters based on similarity. The next step should be to determine each area of adjacent pixels that are located in a cluster. In this project, we try to segment images with this procedure.
There are two solutions for feature extraction, the first method is a method in which the feature vector is determined according to the algorithm designer, while in the second method, learning-based techniques are used to extract the feature vector. Here, we are going to use the second method. Based on this, we are going to use a pre-trained wideResNet-50_2 network for this purpose. We will use the coefficients of this network pre-trained on the ImageNet. The primary layers of this network are responsible for feature extraction.
In this project, we will select only the first six layers of this network and its output will be considered as feature vectors. In the output of this layer, the image size is reduced to one eighth and has 512 channels. This network receives a 224 x 224 input image, so at the output of layer 6, you will have a 28 x 28 x 512 3D matrix (28 is one eighth of 224). That is, we will have 784 512-dimensional vectors = 28x28. Due to the reduction of image dimensions, each of these vectors represents and expresses an 8 x 8 matrix of the input image.

Find the details in MV5-Tabatabaei.pdf.