The Application of FLICSP Algorithm Based on Multi-Feature Fusion in Image Saliency Detection

Abstract

To raise the performance of the current image saliency detection method and promote the extraction effect of salient regions in complex images, from the perspective of multi feature fusion, the image saliency detection method is studied. Corresponding saliency maps is obtained from color and texture channels respectively, and the final saliency map is obtained through linear fusion. On the basis of deep prior information, an improved deep convolutional neural network is used to extract relevant feature vectors such as image pixels, and the nearest neighbor classifier is trained to determine the ownership between pixels and image regions. Zero sample learning iterations are performed to improve the effectiveness of image saliency detection. On this basis, the combination of the above two methods is carried out to obtain the final image saliency detection method. Experimental design is conducted and Matlab software is selected, with a learning rate of 0.001 and 100 iterations. The results showed that compared to other algorithms, the improved algorithm had better detection performance. The detection error of this algorithm was smaller and the accuracy was higher. Its average absolute error value in the JUDD dataset was the smallest, 0.143, which was 0.34 less than the original algorithm. Its detection accuracy in the PASCAL dataset was the highest at 0.786. The missed detection rate and error detection rate of the improved algorithm were both higher than those of the original algorithm, with a missed detection rate of 4.7%, which was 2.5% lower than the original algorithm. Its error detection rate was 5.1%, while the original algorithm’s missed detection rate was 3.2%. The improved algorithm resulted in better image quality, with a maximum peak signal-to-noise ratio of 39.45-dB, which was 6.82-dB higher than the original algorithm. The maximum similarity index value was 0.892. Research methods can effectively detect the saliency of complex images.

Keywords

• FLIC
• VGG network model
• texture features
• color features
• saliency testing