Fusion CNN Based on Feature Selection for Crime Scene Investigation Image Classification

Crime Scene Investigation images have many semantic categories and complex image contents. The Convolution Neural Network (CNN) feature cannot express the uniformity of image content and high dimensional features can lead to redundancy of feature vectors in CNN. In the circumstance it is difficult to use CNN to process crime scene investigation images. To solve the above problems, we propose a fusion CNN algorithm based on feature selection for the classification of crime scene investigation images. In this paper, we build the fusion CNN features to enhance the ability of representation by fusing the convolutional layer with the fully connected layer. Then we select the fusion features with Laplacian score and label mutual information. Finally, we use the obtained features to train Support Vector Machine (SVM) classifier on the Crime Scene Investigation Images Database (CSID). Experiments show that the average classification accuracy of the proposed method can reach 93.67%.