Target Detection Algorithm of Optimized Convolutional Neural Network under Computer Vision

The objectives are to recognize and detect image information such as face contours, to promote the development of convolutional neural network (CNN) algorithms in the field of detection technology, and to expand the application of neural network technology in computer vision, thereby achieving the effective fusion of deep learning and computer vision technology. Based on the problem of low accuracy and meticulousness in traditional contour detection, first, by improving the traditional cross-entropy cost function and Dice coefficient, an adaptive fusion cost function is established. An improved deep reinforcement face contour detection algorithm with mixed training of multi-scale model and adaptive fusion function IDRC-FA-MS is proposed. Second, the AlexNet network is modified by fine-tuning the classifier, and the fully connected layer of CNN is changed to a full convolution layer. A multi-scale face detection algorithm is proposed to analyze the performance indicators related to face detection. Finally, an end-to-end multi-scale face detection algorithm DL-CNN is proposed by connecting the convolution kernel and the convolution layer generated by the feature map to extract the overall features including the detailed features. The results show that after training in the BSDS500 dataset, comparing the IDRC-FA-MS face contour detection algorithm with other algorithms, it is found that the algorithm shows the optimal performance and the highest accuracy reaches 0.99. The average accuracy of the multi-scale face detection algorithm training based on network fine-tuning is 43.4%, the corresponding area of the receiver operating characteristic (ROC) curve is 0.92, and the classification accuracy rate is 99.11%. Also, the detection time is short, and the performance is significantly better than the situation when there is no scale change. Compared with other algorithms, the average accuracy of DL-DNN face detection algorithm is the highest. The corresponding value of continuous score is 0.72 and the corresponding value of discrete score is 0.95. The performance is better than most face detection algorithms.