Resumen
In recent years, deep learning has been widely used in the field of coastal waste detection, with excellent results. However, there are difficulties in coastal waste detection such as, for example, detecting small objects and the low performance of the object detection model. To address these issues, we propose the Multi-Strategy Deconvolution Single Shot Multibox Detector (MS-DSSD) based on DSSD. The method combines feature fusion, dense blocks, and focal loss into a state-of-the-art feed-forward network with an end-to-end training style. In the network, we employ feature fusion to import contextual information to boost the accuracy of small object detection. The dense blocks are constructed by a complex function of three concurrent operations, which can yield better feature descriptions. Then, focal loss is applied to address the class imbalance. Due to the lack of coastal waste datasets, data augmentation is designed to increase the amount of data, prevent overfitting of the model, and speed up convergence. Experimental results show that MS-DSSD513 obtains a higher mAP, of 82.2% and 84.1%, compared to the state-of-the-art object detection algorithms on PASCAL VOC2007 and our coastal waste dataset. The proposed new model is shown to be effective for small object detection and can facilitate the automatic detection of coastal waste management.