Image Salient Object Detection Algorithm Based On Adaptive Multi-Feature Template
Jinping Sun,Enjie Ding,Bo Sun,Lei Chen,Matthew Keith Kerns
DOI: https://doi.org/10.6036/9844
2020-01-01
DYNA
Abstract:Salient object detection affected by area edge blurring and scene complexity has various problems, such as incomplete edge extraction and blurry salient maps. Fusing of multiple salient features improves the detection performance, but an inappropriate fusion algorithm may reduce the results of detection. A salient object detection algorithm based on adaptive multi-feature template was proposed to solve the ineffective fusion of various salient features. First, salient edge features were obtained using Conv1 and Conv2 in the Resnet50 model by combining local edge information with high-level global location information. Second, while some attributes such as texture, color contrast, spatial features, and salient edge features were input into the adaptive multi-feature template, these features were spread to every layer of cellular automata. The final saliency map was obtained by calculating the histogram of the target, background, and entire area of the image and automatically generating weight coefficients of different features according to the intersection of the histogram. Results show that the average absolute error (MAE) of the proposed algorithm is only 0.044, while the comprehensive evaluation index (F-score) reaches 0.899. Thus, this algorithm achieves better accuracy and higher recall rate. The adaptive multi-feature template effectively solves the fusion problem of multiple salient features and can accurately obtain the salient areas of the image. This study provides references for image segmentation, image classification, object tracking, and other fields in computer vision.