Research On Object Detection And Semantic Segmentation Based On Attention Mechanism

In recent years,image object detection and semantic segmentation,as the foundation of computer vision,play an important role in image understanding and analysis.Object detection is mainly concerned with the classification information of objects to be recognized in the image,and the spatial location of these categories of objects is marked.Semantic segmentation refers to the classification of the target object at the pixel level of the image,not only to clarify the category and position of the object,but also to distinguish the specific morphological information of each object.Both methods have excellent application prospects in the field of industrial image detection.Based on this,aiming at the problem of mismatching detection accuracy and speed in blood cell detection of medical images and cloud detection of remote sensing images,combined with the existing research status of deep learning algorithm and the development history of relevant improved technologies,this paper completed the application research in the following two aspects:(1)In the application of blood cell detection,a blood cell detection model based on attention mechanism is proposed to solve the problem of incompatibility between the speed and accuracy of various detection methods.The model is based on YOLO model,and darknet-53 is used as the backbone network,and multi-scale residual enhancement module is added to improve the utilization rate of network feature information.At the same time,an attention-gated plug-in embedding model is designed to integrate more high-quality semantic feature information of upper and lower levels to greatly improve the detection accuracy.Finally,aiming at the problem of insufficient small-target detection ability,the improved focal Loss function based on Focal Loss was adopted.By adding weight values for positive and negative samples,the model could focus more on the samples that were difficult to classify in the training process,thus solving the problem of unbalanced sample categories.(2)In cloud detection applications,a cloud detection model based on attention mechanism is proposed to solve the problem of poor comprehensive performance of u-encoder-decoder network structure in cloud detection tasks.In this model,a new U-shaped encoder-decoder structure is used to jump and connect the shallow and middle information at the encoding end,and an optimization module consisting of block convolution and attention mechanism is embedded at the encoding end.Meanwhile,a context semantic fusion connection is constructed to connect the upper and lower layers of the encoding end and the decoding end.Experimental results show that the proposed model achieves better segmentation accuracy and model parameters.