Research And Application Of SMDH Object Detection Network Based On Improved Modulated Anchoring

Object detection is a multi-tasking technique for detecting the class and location of target objects in an image or video,and it is a crucial research direction in the field of computer vision.Using object detection algorithms to identify target object categories and positions and visualize them accurately can automate people’s work,improve productivity and reduce repetitive tasks.It can also calculate results intelligently,avoiding subjectivity errors in manual recognition and revolutionizing production life for all sectors of society.Laparoscopic instrument detection is one of the most critical applications of object detection in medical image recognition,a technology used to assist and ensure that surgeons can use surgical instruments correctly.Laparoscopic instrument detection automates the detection of the category and location of surgical instruments,helping the surgeon to make a correct and objective judgment of the current surgical outcome,thus assisting the surgeon in ensuring the safety and reliability of the instruments during surgery,reducing the risk of surgery and protecting the patient’s safety.Despite the progress made in the current research of object detection algorithms for practical applications such as laparoscopic surgical instruments,several things could be improved in the inference results of these algorithms,making it difficult to put the studied algorithms into practical applications.One of the main reasons for this phenomenon is that traditional evaluation metrics do not allow for in-depth analysis of the sources of error.Previous work has primarily been analyzed manually and is susceptible to the researcher’s subjectivity.Therefore,this paper introduces an objective,comprehensive evaluation approach based on error analysis using target detection algorithms for laparoscopic surgical instrument detection applications.By evaluating and analyzing three target detection algorithms used for classical laparoscopic surgical instrument detection applications,the main bottlenecks of the existing methods are identified as classification errors,localization errors and missing detection errors.The paper then proposes a modified modifiable anchor frame-based SMDH detection network to address the bottleneck from three perspectives: to address the problem that insufficient extraction capability of the feature extractor is one of the sources of missed detection errors,the paper proposes a method to enhance the extraction capability of the feature extraction network using the Sim AM attention mechanism to infer neuron weights from a three-dimensional perspective,which is more accurate than traditional attention mechanism calculations and does not introduce additional parameter calculations.Furthermore,it does not introduce additional parameters to the computation.In order to enhance the global capacity,local priori and positional perception of the detection head,this paper proposes the MConv structure,which is implemented using convolutional layers to augment the original fully connected layer approach and reconstruct the convolutional parameters at inference time to reduce the computational effort.To address the problem of mutual interference between classification and localization caused by the original single-headed network design,a double-headed detection head with a fully connected layer as the classification branch and a convolutional layer as the localization branch is proposed to replace the single-headed detection head and a SConv block is used to replace the residual block to reduce the number of computational network parameters.Four percentage points improve the overall m AP of the proposed SMDH network compared with the Baseline method,and the classification error,localization error and missed detection error are effectively suppressed by the comprehensive evaluation method proposed in this paper.The visualization of the inference results also proves the effectiveness of this method.The comprehensive evaluation mechanism based on error analysis proposed in this paper can provide a paradigm for subsequent related research to objectively evaluate the strengths and weaknesses of the model,which can be used to accurately analyze the foothold of object detection algorithms and application research and make the research results such as SMDH detection network more compatible with the use of practical applications.