Hybrid camouflage pattern generation using neural style transfer method

Abstract

Manipulating existing camouflage patterns is a challenging issue in the process of camouflage pattern design. In this article, we present an effective approach based on the neural style transfer method to generate a hybrid camouflage pattern by manipulating two given camouflage patterns. Using a convolutional network trained on image recognition, content and style are represented by the correlations between feature maps in several layers of the network. In this regard, we utilized different commonly used camouflage patterns as content and style images. Then, by performing the style transfer algorithm on selected camouflage patterns, a new hybrid camouflage pattern was generated. The hybrid pattern inherits the appearance features of the content and the style images. Also, the colors of the hybrid pattern were controlled by carefully selecting the input colors. It was concluded that the proposed method is useful for adding or deleting some features of an existing camouflage pattern. In fact, it is a professional tool for pixelating, depixelating, blurring, and gradual coloring of camouflage patterns. Furthermore, we demonstrate the effectiveness of hybrid camouflage patterns using visual assessment. The results of the subjective assessment show that the proposed method is efficient for generating successful camouflage patterns.

Efficient digital camouflage pattern generation algorithm based on improved GA-K-means clustering algorithm

“…Five different camouflage patterns are generated under two backgrounds respectively as shown in Figure 7. [11]; (b) the spot combination method [12]; (c) the FCM algorithm [5]; (d) the neural style transfer method [13]; (e) the proposed algorithm.…”

“…Therefore, the color similarity obtained in different backgrounds does not fluctuate significantly. Regarding texture similarity and edge contour similarity, the texture templates of the watershed algorithm [11], spot combination method [12], and neural style transfer method [13] are essentially fixed. As the texture forms of the backgrounds become more complex, these templates become less adaptive, leading to lower texture and edge contour similarity for these algorithms.…”

Application of Image Style Transfer Based on Normalized Residual Network in Art Design

“…The content and style of a CNN trained on image recognition are represented by the correlation between feature maps. The subjective evaluation results indicate that this method is effective in generating successful camouflage patterns [13]. Tian et al proposed a new layout analysis and style fusion system structure for image generation and style conversion.…”

A camouflage generation algorithm based on modified K-means clustering

“…In fact, we need to identify the areas that need to be camouflaged, and mark them with red rectangles in the forest background (Figure 7b) and the ice snow background (Figure 7d) respectively. For comparison, we adopt three reported camouflage generation algorithms [14][15][16] . Firstly, four different camouflage patterns (Figure 8a-d, Figure 9a-d) are generated in the forest and ice snow background.…”