Image salient object detection algorithm based on adaptive multifeature template

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 inappropri-

ate fusion algorithm may reduce the results of detection. A sa-

lient 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 cal

culating 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 pro-

posed algorithm is only 0.044, while the comprehensive evalua-

tion 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 im

age. This study provides references for image segmentation, image

classification, object tracking, and other fields in computer vision.