Abstract: Aiming at the problem that  the traditional object detection had poor feature extraction ability and low recognition rate for small targets, we proposed an improved object detection algorithm based on YOLOv4, which used the attention improved adaptive spatial feature fusion (AIASFF) strategy to generate a pyramid feature representation, and solved the challenges brought by changes in object detection scale. Through this new data-driven pyramid feature fusion strategy, the accuracy of medium and large targets was improved without affecting small target recognition. It combined attention learning image features with extracted features to improve the accuracy of feature detection. The new loss function was combined with the adaptive spatial feature fusion strategy and the exponential moving average,  the simulation results of multiple experiments on the MS COCO dataset based on YOLOv4 show that the algorithm achieves the best compromise between speed and accuracy. For the MS COCO dataset, mAP reaches 41.5% and AP₅₀ reaches 63.8%, which is 1.1% higher than the original algorithm. The improved algorithm has high robustness to MS COCO dataset, thereby  effectively improving the detection and recognition rate of the targets.

Key words: object detection, convolutional neural network, feature pyramid, attention mechanism