Research On Image Fusion Quality Evaluation

Image fusion technology is the integration of multi-modal with the same scene and complementary information to form images with more complete targets and rich information;it avoids the limitation of single sensor imaging principle,and is widely used in medical surgery,military remote sensing,et al.As an important support for the selection and optimization of fusion algorithms,image fusion quality assessment is one of difficulties in image fusion field.Based on the objective evaluation index of image fusion,this paper carries out the following research work:1.A new objective evaluation method for multi-focus image fusion at feature level is proposed.The measure includes mainly two steps: detecting the source image and merging the image to locate the corner point.Then,the overlap of the corners in the fused image and the source image is calculated to represent the quality of the fused image.An attractive feature of the proposed measure is the construction of similarity in the feature level(corner),which avoids focus detection during the evaluation process.The experimental results show that the proposed measure can obtain satisfactory detection accuracy on the data set.2.Based on the high accuracy and low correlation indicator set,a comprehensive index evaluation algorithm for image fusion is also proposed.A comprehensive measure based on measure sets is proposed in this paper.The construction process only consists of two steps: firstly,construct a candidate indicator set,and each of which has high evaluation accuracy and low correlation;Next,dynamically determine the weights of each indicator on the training image set,and finally construct a comprehensive index.Considering that the gold standard for fused image quality comes from subjective perception,we introduce assessment difficulty to reduce the impact of subjective perception errors when calculating the accuracy of each indicator's assessment.The effectiveness of the proposed indicators is tested by conducting various image experiments.Experimental results show that this method is superior to other traditional methods.3.A 3D ROC model is proposed,and it is also applied to the validation of the image fusion measures.In the field of machine learning,the amount of target information carried by each instance is different,in this paper,the kind of instance information that affects learning performance is considered as one of the key point in performance evaluation process.Furthermore,image fusion metrics are abstracted as classifiers,and the validation of the metrics is actually an evaluation of the classifier.A validation method of image fusion quality metrics is proposed,which extended the receiver operating characteristic(ROC)curve into a three-dimensional space and record it as a3 D ROC histogram.In the histogram,the x-axis and the y-axis are identical to the description of the ROC space weight,labeled as a false positive rate and a true positive rate,respectively,and the z-axis is a quantitative indicator indicating the importance of the information carried by each instance.The volume of 3D ROC histogram(V3RH)is used as a summary index.This method preserves the advantages of robustness under class imbalance and independence of threshold.In addition,it provides an easy way to characterize instances during the evaluation process.In this paper,the detail of two simple histograms 3D ROC histogram and V3 RH was also extended.At the same time,the artificial data sets and actual data sets to experiment to verify the performance of the proposed method are compared,and results show that the method is a reliable measure of classifier performance.4.A new image fusion algorithm based on multi-scale and multi-directional decomposition is proposed.By decomposing the high and low frequency sub-bands of the source image,the algorithm can capture the direction information of the source image,and has a strong ability to retain the details such as texture and edge.According to the characteristics of high and low frequency sub-bands,this paper designs different fusion rules.For low-frequency sub-band,this paper uses weighted summation method,in which the size of the weight is related to the local significance of the current pixel,while,gradient maximum method is applied in high frequency sub-band.In this paper,six comparison algorithms and six objective evaluation indexes are tested through multiple groups of medical image tests.The results show that the algorithm has obvious advantages and can get a clearer fusion image,which is in line with the human visual experience.Next step,designing a more efficient multi-scale representation method will help to realize the real-time performance of the algorithm and solve the time cost problem in this algorithm.