Texture segmentation by the 64×64 CNN chip

Abstract

CNN's fast image processing technology helps us to nm high-speed filtering t a s k for image mhaneement, recognition or segmentation. Texme analysis is a spsifie task, since the whole image is pmcessed massively parallel while we b v e B limited number of tame-specific filming and evaluariaa steps. Former resulu of simulations and recognition remlu of simple CNN chips show that the CNN is an appropriate tmI for this imag-prprocessing task. Now we see what the gray-scale image processor CNN chip at i@ limited memory capability and data-handling/-proeetsing accuracy can complele for multi-texm images. We demonstrate and compare some of OUT earlier CNNrelated texhlre analysis methods. Some methods to improve CNN configuration are proposed.

Image Descreening by GA-CNN-Based Texture Classification

“…But for some specific textures like screening patterns mentioned in this paper, the characteristics proposed earlier are not sufficient to classify and segment these screening texture patterns since there may be various causes for screenings such like different resolutions, sampling rates by documental scanning, printing processes, and so on. In [44], the hardware implementation for texture segmentation was developed, which sped up the research and discovery of texture analysis and other related applications, and meanwhile made texture-specific filtering and evaluation processes facilitated with parallel handling capability and consecutive template training of CNNs. The uses of gray level histograms or other statistical methods in [19], [42]- [44] also modulate the decision procedure after texture classification.…”

“…In [44], the hardware implementation for texture segmentation was developed, which sped up the research and discovery of texture analysis and other related applications, and meanwhile made texture-specific filtering and evaluation processes facilitated with parallel handling capability and consecutive template training of CNNs. The uses of gray level histograms or other statistical methods in [19], [42]- [44] also modulate the decision procedure after texture classification. In addition to filtering textures in some orientation, having a process of cross-correlation between the state and input, and using the halftone-like output of CNN as in [47], this paper turns the statistical analysis of CNN and the classification of screening patterns in advance for descreening on a differently breaking view.…”

A comparison of CNN and LEGION networks

“…Their method has been applied to intensity-and motionbased segmentation. Subsequent work has also addressed texture image segmentation [15].…”

“…Several studies also attempt to implement image segmentation tasks, which generally need a hybrid of analog circuits and digital operations. For example, a recent study implements texture classification and segmentation on a 64x64 chip [15].…”